{"name":"RingTru voice AI provider arena","notes":"Vendor claims are in facts, figures measured by others in independent. Each value has its source URL, read date and a quote from the source. RingTru has run no tests of its own on these providers.","sources":[{"name":"Hugging Face Open ASR Leaderboard","url":"https://huggingface.co/spaces/hf-audio/open_asr_leaderboard","who_runs_it":"Hugging Face with NVIDIA, SpeechBrain, Edinburgh/others (paper arXiv 2510.06961). Vendors can submit models, so some rows are vendor-submitted, but scoring is by the leaderboard's own pipeline.","independent":true,"what_it_measures":"Average word error rate over its default English sets (AMI-Cleaned, Earnings22-Cleaned-AA-chunked, Gigaspeech-Cleaned, LS Clean, LS Other, SPGISpeech, URGENT24, Voice Arena Monsoon, Voxpopuli-AA-Cleaned, Private scripted, Private conversational) and RTFx (speed, blank for API-only services).","audio":"Read, meeting, earnings-call, parliamentary and conversational English at 16 kHz. Not phone-codec audio. The 'Private (conversational)' column comes from Appen, DataOcean and Voice Arena sets.","licence_or_terms":"No licence stated on the Space (README front matter has only title, sdk and tags; API cardData has no licence field). The results CSV sits in a private/gated dataset (hf-audio/open-asr-leaderboard-results, english_short_latest.csv returns HTTP 401), so the numbers are only visible through the Space's own rendered table and its public Gradio API. The code repo huggingface/open_asr_leaderboard is Apache-2.0 (code only; GitHub page does not say whether the licence covers results).","terms_quote":"Apache-2.0 license (GitHub sidebar of huggingface/open_asr_leaderboard). Space README: no licence line.","terms_url":"https://github.com/huggingface/open_asr_leaderboard","may_republish":"link only","notes":"Figures below were read on 2026-10-29 from the Space's public Gradio endpoint https://hf-audio-open-asr-leaderboard.hf.space/gradio_api/call/_main_table_update with version_id=latest (leaderboard version 28-10-2026, 'Add nvidia/Nemotron-3.5-Transcribe-30B-A4B.'), the same data the page table shows. 61 rows. Deepgram, Speechmatics, Soniox, Rev, Google STT and Kyutai have no row. Citation to use: Srivastav et al., arXiv 2510.06961."},{"name":"Coval Voice AI Benchmarks (benchmarks.coval.ai)","url":"https://benchmarks.coval.ai/overview","who_runs_it":"Coval (Datawave, Inc.), a voice agent testing company. Not a speech provider, runs its own open-source runner against provider APIs from us-east-1. Coval is a rival testing company to RingTru.","independent":true,"what_it_measures":"STT: word error rate (Whisper EnglishTextNormalizer on both sides), time to first token, time to final segment (from a shared VAD end-of-speech anchor). TTS: time to first audio (includes leading silence) and an intelligibility WER (audio transcribed by OpenAI whisper-1). S2S: voice-to-voice latency from call recordings (clean caller only) plus workflow adherence. Rolling 30 days for STT and TTS, 7 days for S2S, re-measured daily.","audio":"STT datasets: WildASR clean (284 FLEURS clips), accents (845 clips), noise gaps, reverb, far-field, clipping and phone codec (284 clips each through telephony compression), plus PipeCat production turns (897 conversational clips from pipecat stt-benchmark-data). All clips loudness-normalised to -20 dBFS RMS and SHA-256 locked. TTS: 30 fixed short text prompts. S2S: multi-turn simulated calls.","licence_or_terms":"Benchmark runner repo coval-ai/benchmarks is Apache-2.0 (page does not say whether the licence covers published results or datasets; says 'Every number on the site comes from this repository'). The benchmarks site has no stated reuse or attribution terms; footer reads (c) 2026 Coval, Inc. Coval's Terms of Service (last updated May 18, 2026) are written for customers of its Services and say nothing about the public benchmark pages, but clause 8 bars customers from scraping 'any page, data, or portion of or relating to the Services' and clause 7 reserves all rights not expressly granted.","terms_quote":"\"crawl,\" \"scrape,\" or \"spider\" any page, data, or portion of or relating to the Services (or any information, data or content made available through the Services), whether through use of manual or automated means.","terms_url":"https://coval.ai/terms-of-service","may_republish":"link only","notes":"No explicit permission found to republish figures. Safe default: link and cite 'Coval Voice AI Benchmarks, read 2026-10-29' with each figure, or ask Coval in writing. Pages read: /overview, /stt, /s2s, /benchmarks/word-error-rate, /benchmarks/time-to-first-audio, /benchmarks/voice-to-voice-latency. Figures are live rolling averages and will change. Source of the runner licence: https://github.com/coval-ai/benchmarks (quote: \"Licensed under Apache-2.0\"). Coval's server-side location (us-east-1) adds a network round trip to every latency figure."},{"name":"Artificial Analysis Speech to Text","url":"https://artificialanalysis.ai/speech-to-text","who_runs_it":"Artificial Analysis, Inc., an independent benchmarking company (page title: 'Speech to Text (ASR) Providers Leaderboard & Comparison | Artificial Analysis')","independent":true,"what_it_measures":"STT accuracy (AA-WER index), speed and price per provider, methodology at /methodology/speech-to-text","audio":"Not copied. Methodology page not reviewed.","licence_or_terms":"Site licence is personal and noncommercial only; no republishing, no commercial exploitation of site content. Free API requires attribution but is not offered for STT. No numbers copied.","terms_quote":"Company grants you a non-transferable, non-exclusive, revocable, limited license to use and access the Site solely for your own personal, noncommercial use.","terms_url":"https://artificialanalysis.ai/docs/legal/Terms-of-Use.pdf","may_republish":"link only","notes":"Second clause (2.2(d)): \"except as expressly stated herein, no part of the Site may be copied, reproduced, distributed, republished, downloaded, displayed, posted or transmitted in any form or by any means.\" The site also links separate 'Data Platform Terms' (PDF on artificialanalysiscdn.com) that were not read. Free API page (https://artificialanalysis.ai/api-reference) says \"Attribution is required for all use of our free API\" and lists LLM, image, video and text-to-speech endpoints (/data/media/text-to-speech), no STT or speech-to-speech endpoint listed. Link only."},{"name":"Artificial Analysis Text to Speech","url":"https://artificialanalysis.ai/text-to-speech","who_runs_it":"Artificial Analysis, Inc. (page title: 'Best Provider Voice Text to Speech (TTS) Models | Artificial Analysis')","independent":true,"what_it_measures":"TTS quality via its Speech Arena (blind votes), latency and price. No numbers copied.","audio":"Not copied.","licence_or_terms":"Same site terms as above. Free API has a text-to-speech endpoint with required attribution, but its permitted uses are not defined on the page and commercial/redistribution use is not addressed.","terms_quote":"you shall not license, sell, rent, lease, transfer, assign, distribute, host, or otherwise commercially exploit the Site, whether in whole or in part, or any content displayed on the Site","terms_url":"https://artificialanalysis.ai/docs/legal/Terms-of-Use.pdf","may_republish":"link only","notes":"API reference page: https://artificialanalysis.ai/api-reference . Quote there: \"Attribution is required for all use of our free API.\""},{"name":"Artificial Analysis Speech to Speech","url":"https://artificialanalysis.ai/speech-to-speech","who_runs_it":"Artificial Analysis, Inc. (page title: 'Speech to Speech Models and Providers Analysis | Artificial Analysis')","independent":true,"what_it_measures":"Speech-to-speech models: credits tau-Voice, Full Duplex Bench and Big Bench Audio, methodology at /methodology/speech-to-speech-benchmarking. No numbers copied.","audio":"Not copied.","licence_or_terms":"Same site terms as above. No S2S endpoint on the free API.","terms_quote":"no part of the Site may be copied, reproduced, distributed, republished, downloaded, displayed, posted or transmitted in any form or by any means","terms_url":"https://artificialanalysis.ai/docs/legal/Terms-of-Use.pdf","may_republish":"link only","notes":"Page confirmed to exist. Link only."},{"name":"Pipecat STT benchmark (pipecat-ai/stt-benchmark)","url":"https://github.com/pipecat-ai/stt-benchmark","who_runs_it":"Pipecat (open-source voice agent framework from Daily). Not a speech provider, semantic WER is judged by Claude Sonnet 5.5 (an LLM judge, so scores depend on that judge).","independent":true,"what_it_measures":"Semantic WER (only errors that would change an LLM agent's understanding, punctuation, contractions and fillers ignored) and time to final segment (TTFS) for streaming STT services.","audio":"1000 samples from pipecat-ai/smart-turn-data-v3.1-train: short conversational voice-agent turns. Ground truth generated with Gemini and human-reviewed. Not stated as phone-codec audio.","licence_or_terms":"BSD 2-Clause (repository). Results are published in the README.","terms_quote":"BSD 2-Clause","terms_url":"https://github.com/pipecat-ai/stt-benchmark","may_republish":"with credit","notes":"README scoring update (September 2026): \"Semantic WER is now judged by Claude Sonnet 5.5, replacing Claude Sonnet 4.5.\" Rows marked with a dagger were scored by the older judge and 'will be re-scored'. README gives no run date for the table. Raw README read from raw.githubusercontent.com/pipecat-ai/stt-benchmark/main/README.md."},{"name":"tau-Voice (arXiv 2603.13686)","url":"https://arxiv.org/abs/2603.13686","who_runs_it":"Soham Ray, Keshav Dhandhania, Victor Barres, Karthik Narasimhan (submitted 14 March 2026). Academic/benchmark authors, not the vendors.","independent":true,"what_it_measures":"Task completion (pass@1) of full-duplex voice agents on tau2-bench style tasks, plus latency, responsiveness, interruption and selectivity. Providers: OpenAI gpt-realtime-1.5, Google gemini-live-2.5-flash-native-audio, xAI grok-voice-agent.","audio":"Simulated callers (TTS via ElevenLabs v3 at 24 kHz) through G.711 mu-law 8 kHz telephony in both conditions. 'Clean': American accent, no noise. 'Realistic': diverse accents, indoor/outdoor noise, burst noise, ~2% frame drops, muffling, interruptions, coughs.","licence_or_terms":"CC BY 4.0 (arXiv page 'view license' link to creativecommons.org/licenses/by/4.0/).","terms_quote":"Creative Commons Attribution 4.0 (CC BY 4.0)","terms_url":"https://arxiv.org/abs/2603.13686","may_republish":"with credit","notes":"Figures read from the PDF https://arxiv.org/pdf/2603.13686 (text extracted with pdftotext). Models are early 2026 versions (gpt-realtime-1.5, Feb 2026, Gemini Live 2.5 Flash native audio, Dec 2025, grok-voice-agent, Dec 2025), now superseded."},{"name":"Mind the Accent Gap: British Accent Robustness in Speech-Driven Financial Voice Assistants (arXiv 2610.06587)","url":"https://arxiv.org/abs/2610.06587","who_runs_it":"Aadam Haq, Oggi Rudovic, Malcolm Chadwick, Jay Rainey, Shucong Zhang, Ricardo Guerrero, Sourav Bhattacharya, Maja Pantic (submitted 5 October 2026, ICASSP 2027 submission). Academic/industry authors, not the vendors.","independent":true,"what_it_measures":"WER by British accent group and downstream tool-calling exact match for ASR models: NVIDIA Nemotron Streaming and Parakeet-TDT-0.6B, Qwen3-ASR-1.7B, OpenAI Whisper Turbo, Distil-Whisper, Cohere Offline. No Deepgram, AssemblyAI, Speechmatics, Google, Azure, ElevenLabs.","audio":"CavaBench: 90 UK participants, 2,919 retained utterances (3.94 h), clean recording room, spoken financial queries, each recording is also evaluated under 6 acoustic groups (clean, noise, overlap, reverb, environment, narrowband telephony). Table 1 averages over all six.","licence_or_terms":"CC BY 4.0 (arXiv page).","terms_quote":"Creative Commons Attribution 4.0 (CC BY 4.0)","terms_url":"https://arxiv.org/abs/2610.06587","may_republish":"with credit","notes":"Figures from PDF https://arxiv.org/pdf/2610.06587 Table 1. Paper states some models are 'evaluated as offline references'. Dataset is internally collected and not described as public."},{"name":"Responsible ASR: Overcoming Challenges of Foundational Models in Narrow-Band and Low-Resource Settings (arXiv 2606.18659)","url":"https://arxiv.org/abs/2606.18659","who_runs_it":"Tejas Godambe, Nutan Choudhary, Sanket Shah, Nagaraj Adiga, Sharath Adavanne (submitted 17 June 2026). Authors build their own in-house ASR, so their in-house numbers are theirs, the Whisper and Google rows are third-party systems they measured.","independent":true,"what_it_measures":"WER of open-source (Whisper-large-v3, NeMo, Data2Vec AQC, MMS, SpeechBrain) and one commercial system (Google Telephony ASR) on Hindi and Indian-accented English call-centre audio. Only one named commercial provider (Google).","audio":"Spontaneous customer-support telephone calls, narrow-band 8 kHz, 50 hours of fine-tuning and test data, agent and customer channels. Indian-accented English and Hindi. Proprietary recordings, not public.","licence_or_terms":"CC BY 4.0 (arXiv page).","terms_quote":"Creative Commons Attribution 4.0 (CC BY 4.0)","terms_url":"https://arxiv.org/abs/2606.18659","may_republish":"with credit","notes":"Table 2 of the PDF https://arxiv.org/pdf/2606.18659. Model versions are not dated beyond 'Whisper-large-v3' and 'Google Telephony'."},{"name":"TTS Arena V2 (blind listening votes)","url":"https://tts-agi-tts-arena-v2.hf.space/leaderboard","who_runs_it":"TTS-AGI community (Hugging Face Space TTS-AGI/TTS-Arena-V2, the leaderboard app and its JSON are served at tts-agi-tts-arena-v2.hf.space). Providers can list models under a separate Provider Agreement.","independent":true,"what_it_measures":"ELO rating from blind head-to-head votes on which voice sounds more human. Rank, tier, ELO, uncertainty, win rate and total votes per model.","audio":"Short synthesised sentences (prompts from sources such as Gutenberg, Wikipedia, arXiv per TTS-AGI/arena-prompts). Human listeners vote, no phone-codec audio.","licence_or_terms":"Arena software is open source under Apache 2.0 per docs.ttsarena.org. No licence or reuse terms found for the leaderboard data or votes. The Space has a LICENSE file whose content was not shown.","terms_quote":"TTS Arena is open source under Apache 2.0.","terms_url":"https://docs.ttsarena.org/","may_republish":"link only","notes":"Numbers read from the JSON the leaderboard page loads: https://tts-agi-tts-arena-v2.hf.space/api/leaderboard on 2026-10-29. The rank field in the JSON is not monotonic in ELO (it looks tier- or confidence-adjusted), so quote both. Rows flagged preliminary have under ~300 votes. Missing from the arena: OpenAI, Azure, Amazon Polly, Google, Rime, Resemble, Orpheus, Sesame, Deepgram. The older Space TTS-AGI/TTS-Arena is a legacy copy."},{"name":"Picovoice speech-to-text-benchmark","url":"https://github.com/Picovoice/speech-to-text-benchmark","who_runs_it":"Picovoice, which sells competing engines (Leopard, Cheetah). Vendor-run.","independent":false,"what_it_measures":"English WER for batch and streaming engines on LibriSpeech test-clean/test-other, TED-LIUM and CommonVoice, plus punctuation error rate, core-hour and word emission latency.","audio":"Read and talk audio, not phone audio. Cloud engines Amazon Transcribe, Azure, Google, Whisper Large V3 and others are tested alongside Picovoice. Undated results.","licence_or_terms":"Apache-2.0 (repository).","terms_quote":"Apache-2.0 license","terms_url":"https://github.com/Picovoice/speech-to-text-benchmark","may_republish":"link only","notes":"VENDOR-RUN. Label any use as 'measured by Picovoice, a competing vendor'. Results undated, no Deepgram, AssemblyAI, Speechmatics rows. TTS: Picovoice's tts-latency-benchmark (https://github.com/Picovoice/tts-latency-benchmark) tests Amazon Polly, Azure, ElevenLabs (streaming input over websocket), OpenAI TTS, Kokoro, Chatterbox-TTS-Turbo, Pocket-TTS and others against Picovoice Orca for first-token-to-speech latency, but its results are published as plot images (README Results section has image links only), so no figures were taken. Picovoice's web docs page picovoice.ai/docs/benchmark/tts-latency returned 404 when opened."},{"name":"mu-bench: speech-to-text on 8 kHz phone calls (arXiv 2609.32082)","url":"https://arxiv.org/abs/2609.32082","who_runs_it":"Andrea Li (UC Berkeley) and Soham Ray (Sierra, which builds customer-service agents but sells no speech-to-text).","independent":true,"what_it_measures":"Utterance error rate (share of caller turns transcribed wrongly) and latency for speech-to-text services, in five languages, on 4,270 utterances from 250 calls.","audio":"Calls made on callers’ own phones in their own surroundings, at 8 kHz mono over the telephony channel.","licence_or_terms":"CC BY 4.0 (arXiv licence link).","terms_quote":"Creative Commons Attribution 4.0 (CC BY 4.0)","terms_url":"https://arxiv.org/abs/2609.32082","may_republish":"with credit"},{"name":"Full-Duplex-Bench-v3 (arXiv 2604.04847)","url":"https://arxiv.org/abs/2604.04847","who_runs_it":"Lin, Chen, Chen and Lee (National Taiwan University). Academic.","independent":true,"what_it_measures":"Tool use by voice agents when people speak with real hesitations: tool choice, argument accuracy, pass@1, turn-taking, latency and interruptions.","audio":"Real recorded human speech with disfluencies.","licence_or_terms":"CC BY-SA 4.0 (arXiv licence link).","terms_quote":"Creative Commons Attribution-ShareAlike 4.0 (CC BY-SA 4.0)","terms_url":"https://arxiv.org/abs/2604.04847","may_republish":"with credit"},{"name":"Full-Duplex-Bench v1.5 (arXiv 2507.23159)","url":"https://arxiv.org/abs/2507.23159","who_runs_it":"Lin et al. (National Taiwan University). Academic.","independent":true,"what_it_measures":"How speech-to-speech models handle overlapping speech: interruptions, backchannels, side talk and background speech.","audio":"Recorded and synthetic overlap scenarios.","licence_or_terms":"CC BY-SA 4.0 (arXiv licence link).","terms_quote":"Creative Commons Attribution-ShareAlike 4.0 (CC BY-SA 4.0)","terms_url":"https://arxiv.org/abs/2507.23159","may_republish":"with credit"},{"name":"EVA-Bench (arXiv 2605.13841)","url":"https://arxiv.org/abs/2605.13841","who_runs_it":"ServiceNow AI Research. Not a seller of voice models.","independent":true,"what_it_measures":"Task accuracy (EVA-A) and caller experience (EVA-X) of end-to-end voice agents over 213 scenarios, 5 runs each.","audio":"Clean audio.","licence_or_terms":"CC BY 4.0 (arXiv licence link).","terms_quote":"Creative Commons Attribution 4.0 (CC BY 4.0)","terms_url":"https://arxiv.org/abs/2605.13841","may_republish":"with credit"},{"name":"VoiceBench (arXiv 2410.17196)","url":"https://arxiv.org/abs/2410.17196","who_runs_it":"Chen et al. Academic (October 2024).","independent":true,"what_it_measures":"Spoken question answering and instruction following for voice assistants, overall score out of 100.","audio":"Spoken instructions, clean.","licence_or_terms":"CC BY 4.0 (arXiv licence link).","terms_quote":"Creative Commons Attribution 4.0 (CC BY 4.0)","terms_url":"https://arxiv.org/abs/2410.17196","may_republish":"with credit"},{"name":"T5Gemma-TTS technical report (arXiv 2604.01760)","url":"https://arxiv.org/abs/2604.01760","who_runs_it":"Arata and Kurihara, who build their own open text-to-speech model and ran others as baselines. Not a seller in this arena.","independent":true,"what_it_measures":"Naturalness (UTMOS22, a predicted listener score out of 5), word error rate and speaker similarity on LibriSpeech English.","audio":"Synthesised speech, clean.","licence_or_terms":"CC BY 4.0 (arXiv licence link).","terms_quote":"Creative Commons Attribution 4.0 (CC BY 4.0)","terms_url":"https://arxiv.org/abs/2604.01760","may_republish":"with credit"},{"name":"Berkeley Function Calling Leaderboard (BFCL v4)","url":"https://gorilla.cs.berkeley.edu/leaderboard.html","who_runs_it":"UC Berkeley Gorilla project. Academic.","independent":true,"what_it_measures":"How often a model calls the right tool with the right arguments, run by BFCL on its own hosting.","audio":"Text only.","licence_or_terms":"Apache-2.0 (statement in the repo README covering the leaderboard statistics).","terms_quote":"All the leaderboard statistics, and data used to train the models are released under Apache 2.0.","terms_url":"https://github.com/ShishirPatil/gorilla/blob/main/berkeley-function-call-leaderboard/README.md","may_republish":"with credit"}],"providers":[{"id":"vapi","company":"Vapi","product":"Vapi","layer":"platform","homepage":"https://vapi.ai/","summary":"Developer platform that runs a voice agent for you: it joins speech-to-text, a language model and text-to-speech, and charges a hosting fee per minute on top of what those providers charge.","open_weights":false,"facts":{"price":{"value":"$0.05 per minute hosting fee","amount":0.05,"currency":"USD","per":"call_minute","tier":"pay as you go, No Success Package (pre-paid credits)","source":"https://vapi.ai/pricing","read":"2026-10-29","quote":"Usage is your Vapi hosting fee ($0.05/min) plus the model costs for each minute your agents run.","note":"Paid Pro package adds 10% of the hosting fee with a $999/mo minimum (page: '10% of Vapi hosting fee $999/mo minimum'). Premier is contact sales."},"price_note":{"value":"Excludes speech-to-text, language model, text-to-speech and phone-network (carrier) charges, those are billed at provider cost.","source":"https://vapi.ai/pricing","read":"2026-10-29","quote":"Models pass-through at cost","note":"Pricing page calculator for 1,000 minutes lists Hosting Fee $50 separately from Transcriber (Deepgram), Intelligence Model (OpenAI) and Voice (ElevenLabs). Transport fees are 'Charged by provider, not Vapi' (Twilio inbound $0.008/min, outbound $0.014/min on the page), 'Vapi Telephony / SIP' is listed as Free."},"latency":{"value":"Typically around 800 ms end to end","source":"https://docs.vapi.ai/faq","read":"2026-10-29","quote":"Yes, Vapi is designed to achieve low latency, typically around 800 milliseconds for end-to-end voice processing.","note":"Vendor claim from the docs FAQ. The vapi.ai home page separately says '<500ms' as an average (stat) and 'sub-500ms latency'. Vapi's own latency page says the displayed figure adds median transcriber, model and voice latency from production calls and excludes endpointing and transport, so it is a rough estimate."},"uk_numbers":{"value":"No: Vapi's own numbers are US only. A UK number has to be bought elsewhere (Twilio, Telnyx, DIDWW) and imported, or reached over a SIP trunk.","source":"https://docs.vapi.ai/free-telephony","read":"2026-10-29","quote":"Free Vapi numbers use US area codes, and only inbound calls are supported. Import a number from another provider for international use.","note":"Docs page 'Phone Calling' says the same: 'Vapi provides free US phone numbers for inbound calls. For outbound, international, or custom use, import your own numbers from a provider.'"},"bring_your_own":{"value":"Yes: any part of the speech, model and voice pipeline can be swapped, including your own hosted model.","source":"https://docs.vapi.ai/faq","read":"2026-10-29","quote":"You can bring your own custom models for any part of the pipeline.","note":"Own provider keys can be added, or a Custom LLM URL used for models hosted elsewhere."},"telephony":{"value":"No UK numbers sold. SIP trunking supported (bring your own trunk), numbers can be imported from Twilio, Telnyx or DIDWW.","source":"https://docs.vapi.ai/phone-calling","read":"2026-10-29","quote":"Import your own number (for international or custom use)","note":"SIP docs: 'To place outbound calls, import a phone number from a supported provider or connect your own SIP trunk. Use a SIP trunk for international destinations.' Vapi SIP numbers are inbound only. Pricing page lists 'Dedicated SIP available' on Premier."},"data_region":{"value":"EU region (separate isolated environment, endpoints api.eu.vapi.ai and sip.eu.vapi.ai) but only on the Premier package, no UK region found.","source":"https://docs.vapi.ai/security-and-privacy/eu-region","read":"2026-10-29","quote":"Vapi's EU region is available only with the Premier Success Package.","note":"The page does not name a data centre or mention the UK. Premier is 'Contact Sales' on the pricing page."},"compliance":{"value":"SOC 2 Type II attested, GDPR compliant, HIPAA compliant (HIPAA is a paid add-on). ISO 27001 not claimed on the pages checked.","source":"https://docs.vapi.ai/faq","read":"2026-10-29","quote":"HIPAA compliant for healthcare applications SOC 2 Type II-attested GDPR compliant for handling EU data","note":"The FAQ lists these as three bullet points. Pricing page: HIPAA compliance add-on is $2,000 per month ('HIPAA eligible data handling. Standard Data Processing Agreement and Business Associate Agreement'). Docs SOC 2 page: 'Vapi, as a Voice Agent Platform, is a SOC 2 Type II-attested service organization'. HIPAA mode and zero data retention are mutually exclusive per the HIPAA docs page."}},"terms":{"verdict":"restricts other use","source":"https://vapi.ai/terms-of-service","read":"2026-10-29","quote":"try to detect, scan, or test any vulnerabilities of the Services or any of our or our Provider's system or network"}},{"id":"retell-ai","company":"Retell AI","product":"Retell AI","layer":"platform","homepage":"https://www.retellai.com/","summary":"Hosted platform for building phone voice agents, priced as one per-minute total made of a Retell voice fee plus separate model and voice lines.","open_weights":false,"facts":{"price":{"value":"$0.055 per minute (Retell Voice Infra)","amount":0.055,"currency":"USD","per":"call_minute","tier":"pay as you go","source":"https://www.retellai.com/pricing","read":"2026-10-29","quote":"Retell Voice Infra $0.055/minute","note":"Page headline for the all-in total is '$0.07-$0.31 / min for AI Voice Agents'. The page also says 'No platform fees, no feature gating', so the Voice Infra line is the closest thing to Retell's own per-minute fee. The page does not say in words what Voice Infra includes (it sits alongside separate LLM and TTS lines)."},"price_note":{"value":"Separate lines for the language model (example $0.04/min, varies by model) and text-to-speech ($0.015/min for Retell, Minimax, Fish, Cartesia, OpenAI voices, ElevenLabs $0.040 and up). Phone: 'No Charge for Sip Trunking/Custom Telephony', Retell-sold numbers carry their own fees.","source":"https://www.retellai.com/pricing","read":"2026-10-29","quote":"Cost Per Minute $ 0.11 LLM Cost $ 0.04 /min Retell Voice Infra $ 0.055 /min TTS Cost $ 0.015 /min Telephony Cost $ 0.00 /min","note":"Calculator example on the page sums to $0.11/min. Add-ons billed on top (e.g. Knowledge Base, PII Removal). Silence and hold time are billed, calls are billed to the nearest second."},"latency":{"value":"About 600 ms","source":"https://www.retellai.com/","read":"2026-10-29","quote":"With ~600ms latency, conversations stay smooth and fluent.","note":"Vendor claim, no stated method. The same page says 'Independent benchmarks confirm Retell as the leader in responsiveness' but names no benchmark for latency."},"uk_numbers":{"value":"No: Retell sells US and Canada numbers only. A UK number has to come from your own provider (for example Twilio) via custom telephony.","source":"https://docs.retellai.com/deploy/purchase-number","read":"2026-10-29","quote":"Currently we only support purchase of US and Canada numbers","note":"Docs conflict slightly: the International calling page says 'Retell also offers international numbers to purchase' and lists UK outbound at $0.10/min on Twilio, but the purchase page limits buying to US and Canada (the sentence continues 'and support making calls to 15 countries'). The pricing page lists 'United Kingdom (Twilio)' under call rates at $0.015/min, which differs from the $0.10 on the International calling page."},"bring_your_own":{"value":"Language model: yes (your own server over WebSocket). Speech and voices: choose from Retell's listed providers, no own-key option found.","source":"https://docs.retellai.com/integrate-llm/overview","read":"2026-10-29","quote":"Run your own LLM behind a Retell voice agent.","note":"Docs: with a custom LLM 'Retell handles telephony, transcription, and turn-taking while your WebSocket server generates each reply.' No bring-your-own speech-to-text or text-to-speech option was found on the pages checked."},"telephony":{"value":"SIP trunking supported, no charge for SIP trunking or custom telephony.","source":"https://docs.retellai.com/deploy/custom-telephony","read":"2026-10-29","quote":"Connect your telephony provider to Retell AI with SIP trunking or SIP URI dialing.","note":"Guides for Twilio, Telnyx, Vonage plus contact-centre platforms. Pricing page: 'No Charge for Sip Trunking/Custom Telephony'."},"data_region":{"value":"No EU or UK region: Retell says it does not operate services within the EU.","source":"https://docs.retellai.com/general/compliance","read":"2026-10-29","quote":"However, please note that we do not currently operate services within the European Union.","note":"Enterprise on-prem is mentioned on the home page ('within your own infrastructure to meet strict security and data residency requirements') with no region detail. Retell does offer a self-sign DPA with EU Standard Contractual Clauses."},"compliance":{"value":"HIPAA, GDPR, SOC 2 Type 1 and Type 2. The home page also claims ISO 27001, the docs compliance page does not.","source":"https://docs.retellai.com/general/compliance","read":"2026-10-29","quote":"we are proud to be **HIPAA compliant, GDPR compliant, SOC 2 Type 1 & Type 2 certified**","note":"Home page quote: 'Fully compliant with HIPAA, SOC2 Type II, GDPR and ISO 27001.' GDPR claim rests on AWS's DPA ('by utilizing Amazon Web Services (AWS)'). BAA and DPA are self-signable at no extra fee per the docs."}},"terms":{"verdict":"restricts publishing","source":"https://www.retellai.com/legal/terms-of-service","read":"2026-10-29","quote":"(d) use the Services for benchmarking or competitive analysis without Retell AI’s prior written consent"}},{"id":"elevenlabs-agents","company":"ElevenLabs","product":"ElevenAgents (ElevenLabs Agents, formerly Conversational AI)","layer":"platform","homepage":"https://elevenlabs.io/agents","summary":"ElevenLabs' hosted voice agent platform, the per-minute fee covers its speech-to-text and text-to-speech, with the language model billed on top.","open_weights":false,"facts":{"price":{"value":"$0.08 per additional call minute (plans include a monthly allowance of minutes)","amount":0.08,"currency":"USD","per":"call_minute","tier":"after the plan's included minutes; same $0.080 on every paid plan; burst $0.16","source":"https://elevenlabs.io/pricing/agents","read":"2026-10-29","quote":"Additional call minutes cost $0.08 per minute","note":"Plans: Free 15 min, Starter $6/mo 75 min, Creator $22 275 min, Pro $99 1,238 min, Scale $299 3,738 min, Business $990 12,375 min. Burst pricing $0.16/min when above the concurrency limit. The product is branded 'ElevenAgents' on the pricing page."},"price_note":{"value":"Language model usage is billed separately on top. ElevenLabs charges nothing extra for telephony, but any external provider (Twilio, SIP trunk) bills you directly. The page lists Text to Speech, Speech to Text, Realtime, Knowledge Bases RAG and Telephony as included features.","source":"https://elevenlabs.io/pricing/agents","read":"2026-10-29","quote":"LLM usage is billed separately on top, based on the model you choose. ElevenLabs does not charge extra for telephony","note":"Page table headings: 'External Providers LLM Telephony Provider' with 'Based on usage and varies by model. At cost'. Text messages cost $0.003 each."},"latency":{"value":"620 ms time to first audio for the Realtime v1 Mini speech-to-speech model (570 ms for Realtime v1 Max), no figure for the standard speech-to-text, model, text-to-speech pipeline","source":"https://elevenlabs.io/pricing/api","read":"2026-10-29","quote":"96% on conversational dynamics, 620ms for time to first audio","note":"Vendor claim about the Realtime model only, not the whole platform, the Max model line reads '99% on speech reasoning, 570ms for time to first audio'. The agents pricing page makes no latency claim."},"uk_numbers":{"value":null,"source":"https://elevenlabs.io/docs/help-center/product/eleven-agents/can-i-connect-a-non-twilio-phone-to-eleven-agents","read":"2026-10-29","quote":"Yes — you can connect non-Twilio numbers using SIP trunking.","note":"No ElevenLabs-sold phone numbers are mentioned on the pricing or docs pages checked, numbers come from Twilio (native integration) or any SIP trunk provider, so a UK number from either would work but ElevenLabs does not publish UK number sales."},"bring_your_own":{"value":"Language model: yes (custom LLM through an endpoint that follows the OpenAI format). Speech is ElevenLabs' own.","source":"https://elevenlabs.io/docs/eleven-agents/customization/llm/custom-llm","read":"2026-10-29","quote":"Custom LLM allows you to connect your conversations to your own LLM via an external endpoint.","note":"Docs say the server must match OpenAI's Chat Completions or Responses request format."},"telephony":{"value":"SIP trunking and Twilio native integration, other carriers (Telnyx, Vonage, Sinch, Plivo, Bandwidth) documented.","source":"https://elevenlabs.io/docs/eleven-agents/phone-numbers/sip-trunking","read":"2026-10-29","quote":"Connect your existing phone system with ElevenLabs Agents using SIP trunking","note":"SIP docs say TLS transport and SRTP are supported. Static-IP SIP in isolated regions is available to Enterprise accounts."},"data_region":{"value":"Data residency in EU, India and Singapore, Enterprise only, no UK region.","source":"https://elevenlabs.io/docs/overview/administration/data-residency","read":"2026-10-29","quote":"Data residency is an Enterprise feature.","note":"Same page: 'ElevenLabs has released additional storage locations in the EU, India, and Singapore'. Storage is in the chosen location but 'processing may nevertheless occur outside of the selected location', EU processing can be restricted with Zero Retention Mode and the API."},"compliance":{"value":"SOC 2 Type 2, ISO 27001 (plus ISO 27017, 27018, 27701, 42001), PCI DSS Level 1, HIPAA (attestation), GDPR (audit). HIPAA BAA only on Enterprise.","source":"https://start.elevenlabs.io/icps","read":"2026-10-29","quote":"ElevenLabs maintains SOC 2 Type 2, ISO 27001, ISO 27017, ISO 27018, ISO 27701, ISO 42001, PCI DSS Level 1, HIPAA (attestation), GDPR (audit), and additional regional certifications.","note":"Marketing FAQ says these apply across ElevenAgents, ElevenCreative and ElevenAPI. Docs HIPAA page: 'ElevenLabs Agents is one of ElevenLabs' HIPAA-eligible services' and 'Execution of a BAA ... is only available for Enterprise tier subscriptions.'"}},"terms":{"verdict":"restricts other use","source":"https://elevenlabs.io/terms-of-use","read":"2026-10-29","quote":"Using our Services or Output to research or develop products or services that compete with ours"}},{"id":"bland-ai","company":"Bland AI","product":"Bland","layer":"platform","homepage":"https://www.bland.ai/","summary":"Voice agent platform that runs its own speech, model and voice stack and charges one all-in per-minute rate for talk time, with phone charges separate.","open_weights":false,"facts":{"price":{"value":"$0.14 per minute talk time (Start plan), $0.12 per minute on Build with a $299/month platform fee","amount":0.14,"currency":"USD","per":"call_minute","tier":"Start (self-serve, $0 platform fee); Build is $0.12/min plus $299/month","source":"https://www.bland.ai/pricing","read":"2026-10-29","quote":"$0.14 /min $0 platform fee","note":"Transfer time is billed at $0.05/min on Start and $0.04/min on Build. Enterprise is custom. Daily cap 100 calls on Start."},"price_note":{"value":"Rate includes the language model, speech-to-text and text-to-speech. Phone (telephony) is billed separately, on your own carrier or Bland's at pass-through cost.","source":"https://www.bland.ai/pricing","read":"2026-10-29","quote":"Telephony is billed separately, on your own carrier or Bland's at pass-through cost.","note":"Page: 'Included in your per-minute rate LLM No token charges STT Real-time transcription TTS Premium voices'. Bland's docs overview (llms.txt) says per-minute pricing 'covers LLM, TTS, STT, and telephony in a single line item', which conflicts with the pricing page, the pricing page is used here."},"latency":{"value":"Under 400 ms response latency","source":"https://docs.bland.ai/llms.txt","read":"2026-10-29","quote":"The platform delivers sub-400ms response latency","note":"Vendor claim in the docs overview, no method given. The pricing page and home page give no latency figure."},"uk_numbers":{"value":null,"source":"https://docs.bland.ai/api-v1/post/inbound-purchase","read":"2026-10-29","quote":"Options: \"US\" or \"CA\" for Canada. For others, please contact support.","note":"The purchase-number API lists only US and Canada. UK numbers could be imported through your own Twilio account ('Connect your own Twilio account to Bland, easily bring over your existing phone numbers') but Bland does not publish UK number sales."},"bring_your_own":{"value":"No for models and voices: Bland runs its own full stack. Yes for phone service (your own Twilio account or SIP trunk).","source":"https://docs.bland.ai/llms.txt","read":"2026-10-29","quote":"Bland owns the full voice AI stack: telephony, speech-to-text, language model, text-to-speech, and pathway orchestration.","note":"Pricing page: 'Bland integrates with your existing Twilio account, with SIP trunks from any contact-center or telephony provider', 'BYOT customers do not pay transfer fees.'"},"telephony":{"value":"SIP trunks supported (carrier of your choice), Bland also sells US and Canada numbers at $15/month.","source":"https://docs.bland.ai/agents/sip","read":"2026-10-29","quote":"Connect your own carrier to Bland: create trunks, attach numbers and DIDs, port numbers in, and route calls in both directions.","note":"Purchase Phone Number API: '($15/mo. subscription using your stored payment method)'."},"data_region":{"value":"Data residency is an Enterprise-only feature, no regions are named. On-prem or VPC deployment also listed for Enterprise.","source":"https://www.bland.ai/pricing","read":"2026-10-29","quote":"On-prem / VPC available Forward deployed engineer available BAA, SSO, data residency","note":"Comparison table lists 'Data Residency' as a feature row only. No UK or EU region named."},"compliance":{"value":"SOC 2 Type I and II, HIPAA (BAA), GDPR (DPA), PCI DSS self-attested, FedRAMP 20x Class A, ISO 27001","source":"https://www.bland.ai/pricing","read":"2026-10-29","quote":"SOC 2 Type I and Type II (independently audited), HIPAA compliant with a signed BAA, GDPR compliant with a DPA, and PCI DSS self-attested. Bland is also FedRAMP 20x Class A certified and ISO 27001 certified.","note":"Same page: 'Compliance documentation is available under NDA on Enterprise.' BAA is an Enterprise feature in the comparison table."}},"terms":{"verdict":"no clause found","source":"https://www.bland.ai/legal/terms","read":"2026-10-29","quote":"Burdens infrastructure in a way that negatively impacts Service performance, including \"denial of service\" attacks"}},{"id":"synthflow","company":"Synthflow","product":"Synthflow","layer":"platform","homepage":"https://synthflow.ai/","summary":"Voice agent platform that now sells only through enterprise contracts, with its own telephony network.","open_weights":false,"facts":{"price":{"value":null,"amount":null,"currency":"USD","per":"call_minute","tier":"Enterprise only; no per-minute rate published","source":"https://synthflow.ai/pricing","read":"2026-10-29","quote":"Enterprise contracts start at $30,000 annually.","note":"Not published: the pricing page shows one plan, Enterprise, 'scoped around call volume, concurrency, telephony setup, integrations, security needs, and launch support'. No self-serve per-minute list price on the page."},"price_note":{"value":"Enterprise contract includes Synthflow Native Telephony or SIP trunking, SLA, integrations and launch support, exclusions are not stated.","source":"https://synthflow.ai/pricing","read":"2026-10-29","quote":"Synthflow Native Telephony, SIP trunking, or approved enterprise telephony","note":"The page lists what an Enterprise contract covers but does not itemise speech, model or phone costs."},"latency":{"value":"Sub-500 ms (voice agents), telephony network target under 100 ms round trip","source":"https://synthflow.ai/","read":"2026-10-29","quote":"multi-language support, and sub-500ms latency, Synthflow","note":"Vendor claim on the home page. The home page also says 'we keep latency below 100 ms', and the telephony docs set a target 'round-trip time under 100 ms' for audio inside one region, which measures the phone network, not the agent's reply time."},"uk_numbers":{"value":"No: Synthflow sells US, Canadian and Australian numbers only. UK numbers must come from your own Twilio account or a SIP trunk.","source":"https://docs.synthflow.ai/phone-numbers","read":"2026-10-29","quote":"You can buy numbers in the U.S., Canada, and Australia directly through Synthflow.","note":"Docs: 'For local numbers in other regions, import a ...' custom number. Purchased numbers cost $1.50 per month."},"bring_your_own":{"value":null,"source":"https://docs.synthflow.ai/telephony","read":"2026-10-29","quote":"You can buy a number in Synthflow, bring your own Twilio account, or connect your existing carrier or PBX over SIP.","note":"Bring-your-own telephony is documented. No bring-your-own model or speech-provider option was found on the pricing, telephony or docs index pages checked."},"telephony":{"value":"SIP trunking and PBX integration, but only on the Enterprise plan.","source":"https://docs.synthflow.ai/telephony","read":"2026-10-29","quote":"SIP/PBX integration requires an Enterprise plan subscription.","note":"Tier 1 tested providers include Twilio, Telnyx, Vonage and RingCentral RingCX."},"data_region":{"value":"US, EU or legacy Global cluster, chosen when the workspace is created and permanent, no UK cluster.","source":"https://docs.synthflow.ai/customer-region","read":"2026-10-29","quote":"Single-region cluster with all data and processing fully in the European Union.","note":"Global cluster stores data in the EU but processing may run in the US or EU, Synthflow is retiring Global in Q4 2026."},"compliance":{"value":"SOC 2, HIPAA, PCI DSS and GDPR claimed","source":"https://synthflow.ai/","read":"2026-10-29","quote":"Synthflow’s infrastructure is certified across SOC 2, HIPAA, PCI DSS, and GDPR","note":"Home-page marketing text with no report dates or scope. The pricing page shows badges for SOC 2, GDPR, HIPAA and ISO 27001 but ISO 27001 is not stated in page text. Synthflow publishes a HIPAA BAA template in its docs."}},"terms":{"verdict":"no clause found","source":"https://docs.synthflow.ai/terms-conditions","read":"2026-10-29","quote":"except to the extent permitted by applicable law, disassemble, reverse engineer, or decompile the Software."}},{"id":"livekit-cloud-agents","company":"LiveKit","product":"LiveKit Cloud (Agents)","layer":"platform","homepage":"https://livekit.com/","summary":"Hosting and phone plumbing for voice agents you code yourself with the open-source LiveKit Agents framework, you pay per agent minute plus separate model costs.","open_weights":false,"open_source":true,"facts":{"price":{"value":"$0.01 per agent session minute (after plan allowance), plans $0, from $50 and $500 per month","amount":0.01,"currency":"USD","per":"call_minute","tier":"Build ($0/mo, 1,000 min included), Ship (from $50/mo, 5,000 min), Scale ($500/mo, 50,000 min); $0.01 per minute beyond","source":"https://livekit.com/pricing","read":"2026-10-29","quote":"Agent session $ 0.0100 /min","note":"Page: '5,000 minutes included then $0.01 per min'. Agent session minutes count 'Time spent in a session by an agent deployed on LiveKit Cloud'."},"price_note":{"value":"Excludes models: speech-to-text, language model and text-to-speech are charged per minute through LiveKit Inference (examples on page: Deepgram Flux $0.0065/min, GPT-5 mini $0.0011/min, Cartesia Sonic 3 $0.0300/min) or your own provider keys. Phone: US local inbound $0.01/min, third-party SIP minutes $0.004/min on Ship.","source":"https://livekit.com/pricing","read":"2026-10-29","quote":"Telephony $0.0100/min","note":"Page calculator example totals $0.0479/min with Agent session $0.0100, LLM $0.0014, STT $0.0075, TTS $0.0090, observability $0.0100 and telephony $0.0100 (some lines may be optional). Prices differ on the Scale plan."},"latency":{"value":"100 ms global latency (network transport figure, not voice-agent reply time)","source":"https://livekit.com/","read":"2026-10-29","quote":"100ms Global latency","note":"Vendor claim in the home page stats row, it describes the media network, and no end-to-end voice agent response time is published."},"uk_numbers":{"value":"No: LiveKit Phone Numbers are US only and inbound only. A UK number comes through a third-party SIP trunk (Twilio, Telnyx and others).","source":"https://docs.livekit.io/telephony/start/phone-numbers","read":"2026-10-29","quote":"Available only in the US. Support for additional countries is coming in a future release.","note":"Same page: 'LiveKit Phone Numbers currently only supports inbound calling.' Pricing page lists third-party SIP minutes at $0.004/min on Ship."},"bring_your_own":{"value":"Yes: plug in any speech, language model and voice providers through plugins, or use LiveKit Inference with one key.","source":"https://docs.livekit.io/llms.txt","read":"2026-10-29","quote":"with LiveKit Inference and plugins for STT, LLM, TTS, and realtime models","note":"Quote joins the Agents framework description. Agent code is Python or Node.js that you write and deploy."},"telephony":{"value":"SIP trunking supported (inbound and outbound trunks, dispatch rules), US numbers sold by LiveKit, custom SIP domains not available on Build, Ship or Scale.","source":"https://livekit.com/pricing","read":"2026-10-29","quote":"Third-party SIP minutes Inbound and outbound minutes using a third-party SIP trunk","note":"Docs index: 'SIP telephony for agents: phone numbers, inbound and outbound trunks, dispatch rules, provider setup (Twilio, Telnyx, Plivo, and more)'."},"data_region":{"value":"Agents can be deployed to eu-central (Frankfurt) and project data can be kept in the EU, region pinning is a Scale-plan feature. Agent hosting regions are Ashburn, Frankfurt and Mumbai, so no UK agent region.","source":"https://docs.livekit.io/deploy/admin/regions/agent-deployment","read":"2026-10-29","quote":"Deploying to eu-central keeps agent compute in the EU, but this is only one part of configuring a project for EU data residency.","note":"The region list is on https://docs.livekit.io/deploy/admin/regions/endpoints/ (not the page quoted): agent deployment regions are us-east (Ashburn), eu-central (Frankfurt) and ap-south (Mumbai) only. A 'gb' London entry appears there only in the outbound-call destination table, not as an agent region. Pricing page lists 'Region pinning' under Scale."},"compliance":{"value":"SOC 2 Type II, GDPR (DPA available), HIPAA (BAA on Scale and Enterprise). ISO 27001 and PCI DSS in progress.","source":"https://livekit.com/security","read":"2026-10-29","quote":"PCI DSS and ISO 27001 are in progress.","note":"Same page: 'HIPAA BAAs available for Scale and Enterprise customers handling protected health information' and 'GDPR Aligned with EU data protection and privacy law. DPA available'. Pricing page lists 'Security reports / HIPAA' on Scale."}},"terms":{"verdict":"restricts publishing","source":"https://livekit.com/legal/terms-of-service","read":"2026-10-29","quote":"(c) use the Services to build a competing product or service, or benchmark for competitive purposes, without LiveKit's express written consent"}},{"id":"pipecat-cloud","company":"Daily","product":"Pipecat Cloud","layer":"platform","homepage":"https://www.daily.co/products/pipecat-cloud/","summary":"Managed hosting for voice agents written with the open-source Pipecat framework, Daily charges only for the compute minutes and transport, and you pay your speech and model providers directly.","open_weights":false,"open_source":true,"facts":{"price":{"value":"$0.01 per active minute (agent-1x), $0.02 for agent-2x, $0.03 for agent-3x","amount":0.01,"currency":"USD","per":"call_minute","tier":"agent-1x (0.5 vCPU, 1 GB), the profile recommended for voice agents; reserved always-on agents $0.0005/min","source":"https://www.daily.co/pricing/pipecat-cloud/","read":"2026-10-29","quote":"agent-1x Best for voice agents 0.5 vCPU, 1 GB Memory Active $0.01 /min Reserved* $0.0005 /min","note":"Krisp noise filtering is free up to 10k active minutes a month, then $0.0015/min."},"price_note":{"value":"Excludes the language model, speech-to-text and text-to-speech (billed by those providers, enterprise bundled billing by contact). WebRTC transport $0.004 per participant minute. Phone: Daily SIP $0.003 to $0.02/min, Daily PSTN $0.018/min, Twilio telephony billed by Twilio.","source":"https://www.daily.co/pricing/pipecat-cloud/","read":"2026-10-29","quote":"Connect directly with 80+ service providers for LLM, STT, and TTS including Anthropic, Cartesia, Deepgram, ElevenLabs, Google, OpenAI, and more.","note":"Page calculator states it estimates 'Pipecat Cloud (agent hosting, transport, and telephony) and external services not billed via Pipecat Cloud (LLM, STT, and TTS)'. Recording is $0.005/min audio only."},"latency":{"value":null,"source":"https://www.daily.co/products/pipecat-cloud/","read":"2026-10-29","quote":"Multi-region, autoscaling infrastructure.","note":"Not published: no latency figure on the Pipecat Cloud product or pricing pages."},"uk_numbers":{"value":null,"source":"https://docs.pipecat.ai/pipecat/telephony/daily-phone-numbers","read":"2026-10-29","quote":"Purchase and manage phone numbers for Daily's PSTN services to give your Pipecat bots dial-in and dial-out.","note":"Not published: the docs and pricing pages do not say which countries Daily numbers can be bought in (examples are US +1 numbers). Twilio, Telnyx, Plivo and Exotel numbers can be connected by WebSocket instead."},"bring_your_own":{"value":"Yes: you choose every provider for language model, speech-to-text and text-to-speech.","source":"https://www.daily.co/pricing/pipecat-cloud/","read":"2026-10-29","quote":"Integrated Bring your own provider","note":"Page: '80+ service providers for LLM, STT, and TTS ... Billed to provider'. Agent code is yours."},"telephony":{"value":"SIP dial-in and dial-out through Daily ($0.003 to $0.02/min), plus Daily PSTN and carrier WebSockets (Twilio, Telnyx, Plivo, Exotel).","source":"https://www.daily.co/pricing/pipecat-cloud/","read":"2026-10-29","quote":"Daily SIP Dial-in/Dial-out Integrated SIP transport for voice agents $0.003 - $0.02 per minute","note":"Call transfer by SIP Refer costs $0.20 per event on the page."},"data_region":{"value":"Agents can run in Frankfurt (eu-central), US West, US East or Mumbai, no UK region. Custom data residency is an Enterprise feature.","source":"https://docs.pipecat.ai/pipecat-cloud/guides/regions","read":"2026-10-29","quote":"eu-central: Europe (Frankfurt)","note":"Page also offers self-hosted regions in your own cluster through Pipecat Enterprise. Pricing page: Enterprise Security 'SSO integration, custom data residency, and advanced security controls'."},"compliance":{"value":"HIPAA and GDPR claimed for Pipecat Cloud, Daily (the operator) is SOC 2 Type 2 certified. ISO 27001 only for Daily's AWS data centres.","source":"https://www.daily.co/products/pipecat-cloud/","read":"2026-10-29","quote":"HIPAA and GDPR-compliant with enterprise-grade security, data privacy, and compliance.","note":"Daily security page: 'Daily is SOC 2 Type 2 certified', 'Data centers are SOC 1, SOC 2, and ISO 27001 certified', 'we will sign a Business Associates Agreement'. That page covers Daily's video platform generally, not Pipecat Cloud by name."}},"terms":{"verdict":"no clause found","source":"https://www.daily.co/legal/terms-of-service/","read":"2026-10-29","quote":"You shall not (a) use the Services in any unlawful manner (including without limitation in violation of any data, privacy or export control laws), (b) in any manner that interferes with or disrupts the integrity or performance of the Services or their components, or (c) modify, adapt or hack the Services to, or otherwise attempt to gain unauthorized access to, the Services or their related systems or networks."}},{"id":"deepgram-voice-agent","company":"Deepgram","product":"Voice Agent API","layer":"platform","homepage":"https://deepgram.com/product/voice-agent","summary":"One streaming connection that does the listening, thinking and speaking for a voice agent, billed at a single per-minute rate while the connection is open.","open_weights":false,"facts":{"price":{"value":"$0.075 per minute (Standard, pay as you go)","amount":0.075,"currency":"USD","per":"call_minute","tier":"pay as you go, Standard tier; billed on WebSocket connection time","source":"https://deepgram.com/pricing","read":"2026-10-29","quote":"Standard $0.075 / min $0.068 / min","note":"Other pay-as-you-go tiers: Standard with your own text-to-speech $0.065, Custom with your own model $0.065, own model and own text-to-speech $0.050, Advanced $0.163. Growth (prepaid annual) rates are lower, for example $0.068 for Standard. The Growth column is the second figure."},"price_note":{"value":"Standard and Advanced rates include speech-to-text, a managed language model and text-to-speech. Phone numbers and phone-network charges are not part of the rate (Deepgram does not sell numbers).","source":"https://deepgram.com/pricing","read":"2026-10-29","quote":"One rate covering speech-to-text, a managed language model, and text-to-speech. Bring your own LLM or TTS provider for a lower per-minute rate.","note":"Billing note on page: 'Calculated on WebSocket connection time.' Docs: audio can come from 'a phone call over your telephony provider'."},"latency":{"value":null,"source":"https://deepgram.com/product/voice-agent","read":"2026-10-29","quote":"Every conversation turn returns a latency report, broken down by speech-to-text, LLM time to first token, tool time, and text-to-speech.","note":"Not published: no latency figure on the pricing, product or architecture pages, the product returns a per-turn latency report you can read yourself."},"uk_numbers":{"value":"No: Deepgram does not sell phone numbers, a UK number comes from your own carrier (Twilio, Plivo and others).","source":"https://developers.deepgram.com/docs/voice-agent-architecture","read":"2026-10-29","quote":"Audio streams in from any source — a browser, a mobile app, or a phone call over your telephony provider.","note":"Docs list guides for Twilio, Plivo, Genesys Cloud, Amazon Connect and AudioCodes. No number purchase is described."},"bring_your_own":{"value":"Yes: bring your own language model (or custom endpoint) and your own text-to-speech provider. Speech-to-text is Deepgram's.","source":"https://developers.deepgram.com/docs/voice-agent-architecture","read":"2026-10-29","quote":"Use a Deepgram-managed model, bring your own provider, or point at a custom endpoint.","note":"Product page: 'Bring your own LLM and voices or use Deepgram's.'"},"telephony":{"value":null,"source":"https://developers.deepgram.com/docs/voice-agent","read":"2026-10-29","quote":"connect voice agents to phone networks for inbound and outbound calls.","note":"Not published as a built-in feature: no direct SIP trunk or numbers are described, phone calls reach the API through a carrier bridge such as Twilio ('bridges Twilio call audio to the Deepgram Voice Agent API')."},"data_region":{"value":"Global, EU, India or Australia endpoints, a dedicated EU endpoint keeps processing in the EU. No UK endpoint.","source":"https://deepgram.com/product/voice-agent","read":"2026-10-29","quote":"Global, EU, India, or Australia endpoints","note":"Pricing page: 'We offer a dedicated EU endpoint to ensure data processing stays within the European Union (api.eu.deepgram.com).' Growth concurrency is 'up to 45 in Europe, Australia, and India'."},"compliance":{"value":"SOC 2 Type 1 and Type 2, HIPAA (BAA for Enterprise), GDPR, CCPA, PCI","source":"https://deepgram.com/pricing","read":"2026-10-29","quote":"SOC 2 Type 1 and Type 2, HIPAA with a BAA for Enterprise, GDPR with an EU endpoint, CCPA, and PCI.","note":"Company-wide answer on the pricing page FAQ, not specific to the Voice Agent API."}},"terms":{"verdict":"restricts publishing","source":"https://deepgram.com/terms","read":"2026-10-29","quote":"use or access our Services or any Output for competitive purposes, including model training, benchmarking and other competitive analysis, or developing competing models, products or services;"}},{"id":"twilio-conversationrelay","company":"Twilio","product":"ConversationRelay","layer":"platform","homepage":"https://www.twilio.com/en-us/products/conversational-ai","summary":"Twilio handles the call and converts speech to text and text to speech, then passes the words to a server you run, where your own AI model writes the replies.","open_weights":false,"facts":{"price":{"value":"$0.07 per minute","amount":0.07,"currency":"USD","per":"call_minute","tier":"pay as you go, starting price","source":"https://twilio.com/en-us/products/conversational-ai/pricing","read":"2026-10-29","quote":"Conversation Relay $0.07 / minute","note":"Page: 'Conversation Relay uses pay-as-you-go pricing (starting at $0.07 per minute) with voice costs billed separately.' Whether the speech providers' own fees are inside the $0.07 is not stated on the pricing page."},"price_note":{"value":"Excludes the phone call itself (Twilio voice minutes) and your own language model. The pricing page does not say whether speech-to-text and text-to-speech are inside the $0.07.","source":"https://twilio.com/en-us/products/conversational-ai/pricing","read":"2026-10-29","quote":"Voice costs are calculated separately","note":"Twilio's docs say 'Let Twilio handle the heavy lifting of speech recognition, text-to-speech, and voice synthesis' but give no price breakdown. UK voice for reference: local numbers $3.50/mo, receive $0.0100/min (see Twilio UK voice pricing page, current as of August 2026)."},"latency":{"value":null,"source":"https://www.twilio.com/docs/voice/conversationrelay","read":"2026-10-29","quote":"Let Twilio handle the heavy lifting of speech recognition, text-to-speech, and voice synthesis.","note":"Not published: no latency figure on the ConversationRelay docs, TwiML reference or pricing pages."},"uk_numbers":{"value":"Yes: Twilio sells UK numbers (local numbers $3.50 per month, mobile $2.50, toll-free $2.70), and ConversationRelay runs on Twilio voice calls.","source":"https://www.twilio.com/en-us/voice/pricing/gb","read":"2026-10-29","quote":"Local numbers $0.0100 / min + $3.50 / mo","note":"Page says 'Pricing current as of August 2026'. UK numbers need regulatory documents (see carrier layer). Whether ConversationRelay works on every UK number type is not stated."},"bring_your_own":{"value":"Language model: yes (your own WebSocket server writes the replies). Speech: pick from Google or Deepgram for speech-to-text and Google, Amazon or ElevenLabs for text-to-speech, no own-key option found.","source":"https://www.twilio.com/docs/voice/twiml/connect/conversationrelay","read":"2026-10-29","quote":"The provider for TTS. Available choices are Google, Amazon, and ElevenLabs.","note":"The same page gives the speech-to-text choices as Google and Deepgram. Overview: 'Your application uses AI to analyze the text and generate a response.'"},"telephony":{"value":null,"source":"https://www.twilio.com/docs/voice/conversationrelay","read":"2026-10-29","quote":"Conversation Relay overview","note":"Not published for ConversationRelay itself: calls arrive on Twilio numbers or Twilio voice, and Twilio separately sells Elastic SIP trunking, but the ConversationRelay pages do not say whether SIP-trunk calls can use it."},"data_region":{"value":null,"source":"https://www.twilio.com/docs/global-infrastructure/understanding-twilio-regions","read":"2026-10-29","quote":"Twilio's default Region (US1) is located in the eastern United States.","note":"Not published for ConversationRelay: the regions page lists US1, Ireland and Australia in general, warns 'Not all products and features are available in Regions outside of US1', and does not mention ConversationRelay. No UK region."},"compliance":{"value":"Twilio-wide: SOC 2 Type II and Type I, ISO/IEC 27001:2013, HIPAA-eligible products. Not confirmed for ConversationRelay.","source":"https://www.twilio.com/en-us/security","read":"2026-10-29","quote":"Please note the credentials listed do not apply to every product across Twilio, Segment, and SendGrid.","note":"Page lists SOC 2, ISO/IEC 27001, HIPAA, PCI DSS Level 1, GDPR is not listed as a certification there. The ConversationRelay pages do not state which attestations cover it."}},"terms":{"verdict":"no clause found","source":"https://www.twilio.com/en-us/legal/tos","read":"2026-10-29","quote":"Reverse Engineering. Reverse engineering, copying, disassembling, or decompiling the Services."}},{"id":"ultravox","company":"Ultravox (Fixie.ai)","product":"Ultravox Realtime","layer":"platform","homepage":"https://www.ultravox.ai/","summary":"Voice agent platform built around its own speech-native model that listens to audio directly instead of converting it to text first. Its site says Inworld has acquired it.","open_weights":false,"facts":{"price":{"value":"$0.05 per minute (first 30 minutes free)","amount":0.05,"currency":"USD","per":"call_minute","tier":"pay as you go ($0 per month, up to 5 concurrent calls); Pro is $100/month with no hard concurrency cap","source":"https://www.ultravox.ai/pricing","read":"2026-10-29","quote":"First 30 minutes free, $0.05/min after that.","note":"Pro per-minute rate is also $0.05 per minute on the feature table. Text-based Threads are priced per token ($2.00 per 1M uncached input, $15.00 per 1M output)."},"price_note":{"value":"The page does not itemise what the $0.05 covers. Phone connection over SIP costs extra (0.5 cents per minute). Phone numbers are not sold: connect Twilio, Telnyx, Plivo, Exotel or your own SIP trunk. Using your own text-to-speech provider means paying that provider separately.","source":"https://www.ultravox.ai/pricing","read":"2026-10-29","quote":"SIP Pricing .5¢ per minute .48¢ per minute Custom","note":"SIP docs: 'SIP calls incur costs for SIP minutes in addition to the regular cost of call minutes.'"},"latency":{"value":null,"source":"https://www.ultravox.ai/","read":"2026-10-29","quote":"To keep latency low, we manage our full inference stack, so there's no waiting on external LLMs or shared inference pools.","note":"Not published: no millisecond figure on the home or pricing pages, the site shows a speed-versus-intelligence chart with no stated method here."},"uk_numbers":{"value":null,"source":"https://docs.ultravox.ai/telephony/overview","read":"2026-10-29","quote":"Native support for Twilio, Telnyx, Plivo, and Exotel.","note":"Not published: Ultravox does not sell numbers on the pages checked, a UK number would come from one of those carriers or a SIP trunk."},"bring_your_own":{"value":"Text-to-speech: yes (your own ElevenLabs, Cartesia or other provider keys). The language model is Ultravox's own speech-native model.","source":"https://docs.ultravox.ai/voices/bring-your-own","read":"2026-10-29","quote":"Ultravox Realtime allows you to bring your own TTS provider to have your agent sound however you like.","note":"Home page: 'Ultravox is built on open weight models.' No bring-your-own language model option found."},"telephony":{"value":"SIP supported (inbound and outbound, via partner Voximplant), native Twilio, Telnyx, Plivo and Exotel connections.","source":"https://docs.ultravox.ai/telephony/overview","read":"2026-10-29","quote":"Connect your existing phone system through our SIP integration.","note":"SIP pricing 0.5 cents a minute on pay as you go, billed in 6-second steps."},"data_region":{"value":null,"source":"https://www.ultravox.ai/pricing","read":"2026-10-29","quote":"Trust","note":"Not published: pricing, home and telephony docs pages name no data regions. The footer links to a Trust page (trust.delve.co/ultravox) that blocked automated reading behind a Vercel security checkpoint, so it was not read."},"compliance":{"value":null,"source":"https://www.ultravox.ai/pricing","read":"2026-10-29","quote":"Trust","note":"Not published on the pages read: the pricing page only links to a Trust page (trust.delve.co/ultravox), which could not be opened by the tool (Vercel security checkpoint). No SOC 2, HIPAA, GDPR or ISO 27001 statement found elsewhere."}},"terms":{"verdict":"restricts other use","source":"https://www.ultravox.ai/terms","read":"2026-10-29","quote":"Mandatory AI disclosure . Always clearly and prominently disclose to users they are interacting with AI rather than a human."}},{"id":"deepgram-nova-3","company":"Deepgram","product":"Nova-3","layer":"stt","homepage":"https://deepgram.com/","summary":"Deepgram's general-purpose speech-to-text model, used for live streaming and recorded audio.","open_weights":false,"facts":{"price":{"value":"$0.0077 per minute","amount":0.0077,"currency":"USD","per":"audio_minute","tier":"pay as you go, streaming, Nova-3 monolingual (English)","source":"https://deepgram.com/pricing","read":"2026-10-29","quote":"Nova-3 Monolingual $0.0077 / min $0.0065 / min","note":"The second figure on the row is the growth-plan rate. Nova-3 Multilingual is $0.0092 per minute pay as you go. Page also says: \"New projects should start with Nova-3 for general transcription or Flux for real-time voice agents.\""},"latency":{"value":"No latency figure for Nova-3 streaming. Only a relative claim: comparable to Nova-2.","source":"https://deepgram.com/learn/introducing-nova-3-speech-to-text-api","read":"2026-10-29","quote":"In benchmark testing, Nova-3 demonstrated comparable latency to Nova-2, delivering rapid transcription with minimal delay.","note":"No milliseconds published on the pricing page or the Nova-3 launch post."},"accuracy":{"value":"Median word error rate 6.84% on streaming audio (vendor figure)","source":"https://deepgram.com/learn/introducing-nova-3-speech-to-text-api","read":"2026-10-29","quote":"Nova-3 boasts a median WER of 6.84% on real-time audio streams from a dataset representing diverse real-world scenarios.","note":"Vendor's own test set: 2,703 audio files, 81.69 hours, described as nine domains (the post lists ten). Batch median WER 5.26%. Vendor-run, not independent."},"en_gb":{"value":"yes","source":"https://developers.deepgram.com/docs/models-languages-overview","read":"2026-10-29","quote":"English: en, en-US, en-AU, en-CA, en-GB, en-IE, en-IN, en-NZ","note":"Quote is the nova-3-general row of the models table."},"streaming":{"value":"yes","source":"https://developers.deepgram.com/docs/models-languages-overview","read":"2026-10-29","quote":"audio in batch or streaming","note":"Nova-3 is recommended for batch or streaming, it has no built-in turn detection."},"languages":{"value":"Real-time code-switching across 10 languages (English, Spanish, French, German, Hindi, Russian, Portuguese, Japanese, Italian, Dutch)","source":"https://deepgram.com/learn/introducing-nova-3-speech-to-text-api","read":"2026-10-29","quote":"English, Spanish, French, German, Hindi, Russian, Portuguese, Japanese, Italian, and Dutch.","note":"Quote is the 10 real-time code-switching languages. No total language count was found on the pages read, so none is given."},"data_region":{"value":"Dedicated EU endpoint (api.eu.deepgram.com). No UK-specific region listed.","source":"https://deepgram.com/pricing","read":"2026-10-29","quote":"We offer a dedicated EU endpoint to ensure data processing stays within the European Union","note":"FIPS 140-3 images are for self-hosted deployments only."},"compliance":{"value":"SOC 2 Type 1 and Type 2, GDPR, CCPA, PCI, HIPAA BAA for Enterprise customers","source":"https://deepgram.com/pricing","read":"2026-10-29","quote":"SOC 2 Type 1 & Type 2 Certified","note":"Other quotes on the same page: \"We sign Business Associate Agreements (BAA) for Enterprise customers handling electronic Protected Health Information (ePHI).\" and \"Fully compliant with GDPR.\""}},"independent":[{"by":"Pipecat STT benchmark (semantic WER judged by Claude Sonnet 5.5)","metric":"semantic WER, mean across 1000 conversational voice-agent turns (lower is better)","value":"1.32%","source":"https://github.com/pipecat-ai/stt-benchmark","read":"2026-10-29","quote":"Columns: Vendor, Model, Transcripts, Perfect, WER Mean, Pooled WER, TTFS Median, TTFS P95, TTFS P99. Row: | Deepgram | nova-3-general | 99.8% | 77.7% | 1.32% | 1.37% | 247ms | 298ms | 326ms |","note":"Semantic WER counts only errors that change meaning for an LLM agent. Judge is an LLM (Claude), so scores depend on it. README says WER fell about 13% on average after the September 2026 judge change. Streaming services. The source calls the model \"Deepgram nova-3-general\"."},{"by":"Pipecat STT benchmark (semantic WER judged by Claude Sonnet 5.5)","metric":"time to final segment (TTFS), median","value":"247ms","source":"https://github.com/pipecat-ai/stt-benchmark","read":"2026-10-29","quote":"Columns: Vendor, Model, Transcripts, Perfect, WER Mean, Pooled WER, TTFS Median, TTFS P95, TTFS P99. Row: | Deepgram | nova-3-general | 99.8% | 77.7% | 1.32% | 1.37% | 247ms | 298ms | 326ms |","note":"TTFS measured from a shared end-of-speech anchor to the final transcript, run location and date not stated in README. The source calls the model \"Deepgram nova-3-general\"."},{"by":"mu-bench (arXiv 2609.32082)","metric":"English (en-US) utterance error rate on 8 kHz phone calls, batch, lower is better","value":"3.9%","source":"https://arxiv.org/abs/2609.32082","quote":"Deepgram Nova-3 | batch | en-US 3.9 | all 28.0 | p50 107 ms","note":"Table 5. Share of caller turns transcribed with an error. Batch mode, the one mode the paper ran for all four. The \"all\" column covers five languages (en-US, es-MX, tr-TR, vi-VN, zh-CN), where the spread is far wider.","read":"2026-10-30"}],"terms":{"verdict":"restricts publishing","source":"https://deepgram.com/terms","read":"2026-10-29","quote":"use or access our Services or any Output for competitive purposes, including model training, benchmarking and other competitive analysis, or developing competing models, products or services;"}},{"id":"deepgram-flux","company":"Deepgram","product":"Flux","layer":"stt","homepage":"https://deepgram.com/","summary":"Deepgram's streaming model for voice agents, with built-in detection of when the caller has finished speaking.","open_weights":false,"facts":{"price":{"value":"$0.0077 per minute","amount":0.0077,"currency":"USD","per":"audio_minute","tier":"pay as you go, streaming, Flux English","source":"https://deepgram.com/pricing","read":"2026-10-29","quote":"Flux STT English $0.0077 / min $0.0065 / min","note":"Flux Multilingual is $0.0078 per minute pay as you go."},"latency":{"value":"End-of-turn detection: 1 second at p90, 1.5 seconds at p95 (vendor figure)","source":"https://deepgram.com/learn/introducing-flux-conversational-speech-recognition","read":"2026-10-29","quote":"a p90 (p95) latency of 1 second (1.5 seconds)","note":"Measures the delay from the caller stopping to the end-of-turn signal, not word-by-word transcript delay. Docs quickstart separately says end-of-turn detection at roughly 260 ms. The launch post also says Flux cuts agent response latency by 200 to 600 ms against pipeline approaches."},"accuracy":{"value":null,"source":"https://deepgram.com/learn/introducing-flux-conversational-speech-recognition","read":"2026-10-29","quote":"the lowest WER among the models evaluated","note":"Claim only, no WER number and no named dataset ('our comprehensive STT benchmark suite'). Vendor says WER was nearly identical to Nova-3."},"en_gb":{"value":"English, all accents, through one 'en' option, no separate en-GB code","source":"https://developers.deepgram.com/docs/models-languages-overview","read":"2026-10-29","quote":"English (all accents): en","note":"Docs list Flux English (flux-general-en) as one option covering all accents, with no en-GB code. Nova-3 lists en-GB separately."},"streaming":{"value":"yes","source":"https://developers.deepgram.com/docs/models-languages-overview","read":"2026-10-29","quote":"Our latest-generation streaming model unifying best-in-class ASR with model-native turn detection.","note":""},"languages":{"value":"English, plus a multilingual variant covering 10 languages","source":"https://developers.deepgram.com/docs/flux/quickstart","read":"2026-10-29","quote":"extends Flux to 10 languages with optional language_hint biasing","note":""},"data_region":{"value":"Dedicated EU endpoint (api.eu.deepgram.com). No UK-specific region listed.","source":"https://deepgram.com/pricing","read":"2026-10-29","quote":"We offer a dedicated EU endpoint to ensure data processing stays within the European Union","note":"FIPS 140-3 images are for self-hosted deployments only."},"compliance":{"value":"SOC 2 Type 1 and Type 2, GDPR, CCPA, PCI, HIPAA BAA for Enterprise customers","source":"https://deepgram.com/pricing","read":"2026-10-29","quote":"SOC 2 Type 1 & Type 2 Certified","note":"Other quotes on the same page: \"We sign Business Associate Agreements (BAA) for Enterprise customers handling electronic Protected Health Information (ePHI).\" and \"Fully compliant with GDPR.\""}},"terms":{"verdict":"restricts publishing","source":"https://deepgram.com/terms","read":"2026-10-29","quote":"use or access our Services or any Output for competitive purposes, including model training, benchmarking and other competitive analysis, or developing competing models, products or services;"}},{"id":"assemblyai-universal-3-6-pro-realtime","company":"AssemblyAI","product":"Universal-3.6 Pro Realtime","layer":"stt","homepage":"https://www.assemblyai.com/","summary":"AssemblyAI's top live speech-to-text model, marked as its flagship streaming model.","open_weights":false,"facts":{"price":{"value":"$0.45 per hour (about $0.0075 per minute)","amount":0.0075,"currency":"USD","per":"audio_minute","tier":"pay as you go, streaming, per-minute figure derived from the hourly rate","source":"https://www.assemblyai.com/pricing","read":"2026-10-29","quote":"Supports 32 languages with self-correcting speaker labels, voice isolation, and keyterm prompting. $0.45 /hr","note":"Billing docs (https://www.assemblyai.com/docs/billing-and-pricing): \"Streaming Speech-to-Text is billed on the total duration that your WebSocket connection stays open, not on the amount of audio you send.\" So idle open connections cost money. Add-ons (diarization, keyterms, Voice Focus) cost extra."},"latency":{"value":null,"source":"https://www.assemblyai.com/docs/streaming/universal-streaming","read":"2026-10-29","quote":"Fastest","note":"Docs give only relative labels ('Fastest' for Universal-3.6 Pro, 'Fast' for Universal-Streaming), no milliseconds, no definition of what is measured."},"accuracy":{"value":null,"source":"https://www.assemblyai.com/docs/streaming/universal-streaming","read":"2026-10-29","quote":"Best","note":"Docs rate entity accuracy as 'Best' (3.6 Pro) vs 'Okay' (Universal-Streaming). No word error rate or dataset published on that page."},"en_gb":{"value":null,"source":"https://www.assemblyai.com/pricing","read":"2026-10-29","quote":"","note":"No mention of British English or accents on the pricing page or the streaming docs page We read."},"streaming":{"value":"yes","source":"https://www.assemblyai.com/docs/streaming/universal-streaming","read":"2026-10-29","quote":"Our flagship streaming model.","note":"Quote is the docs description of Universal-3.6 Pro."},"languages":{"value":"32 languages","source":"https://www.assemblyai.com/docs/streaming/universal-streaming","read":"2026-10-29","quote":"32 on Universal-3.6 Pro","note":""},"data_region":{"value":"EU data zone endpoint (wss://streaming.eu.assemblyai.com/v3/ws) processes streaming audio and transcripts in the EU, the default endpoint routes to the nearest region, no UK region listed","source":"https://www.assemblyai.com/docs/streaming/endpoints-and-data-zones","read":"2026-10-29","quote":"Data Zone endpoints will process your audio and transcript data in the specified region.","note":"The security page snippet mentioned a Dublin data centre but the page text read did not say Dublin, so it is not recorded."},"compliance":{"value":"SOC 2 Type 2 report, signed BAA available (HIPAA), ISO 27001 and GDPR not confirmed on the pages read","source":"https://www.assemblyai.com/security","read":"2026-10-29","quote":"We maintain a SOC 2 Type 2 report covering security, availability, and confidentiality, audited annually by an independent firm.","note":"Security page also says: \"A signed BAA opts you out automatically.\" The pricing page does not state ISO 27001 or GDPR in its text, so those are not claimed here."}},"independent":[{"by":"Pipecat STT benchmark (semantic WER judged by Claude Sonnet 5.5)","metric":"semantic WER, mean across 1000 conversational voice-agent turns (lower is better)","value":"0.89%","source":"https://github.com/pipecat-ai/stt-benchmark","read":"2026-10-29","quote":"Columns: Vendor, Model, Transcripts, Perfect, WER Mean, Pooled WER, TTFS Median, TTFS P95, TTFS P99. Row: | AssemblyAI | universal-3-6-pro | 99.9% | 88.7% | 0.89% | 0.80% | 307ms | 401ms | 498ms |","note":"Semantic WER counts only errors that change meaning for an LLM agent. Judge is an LLM (Claude), so scores depend on it. README says WER fell about 13% on average after the September 2026 judge change. Streaming services. The source calls the model \"AssemblyAI universal-3-6-pro\"."},{"by":"Pipecat STT benchmark (semantic WER judged by Claude Sonnet 5.5)","metric":"time to final segment (TTFS), median","value":"307ms","source":"https://github.com/pipecat-ai/stt-benchmark","read":"2026-10-29","quote":"Columns: Vendor, Model, Transcripts, Perfect, WER Mean, Pooled WER, TTFS Median, TTFS P95, TTFS P99. Row: | AssemblyAI | universal-3-6-pro | 99.9% | 88.7% | 0.89% | 0.80% | 307ms | 401ms | 498ms |","note":"TTFS measured from a shared end-of-speech anchor to the final transcript, run location and date not stated in README. The source calls the model \"AssemblyAI universal-3-6-pro\"."}],"terms":{"verdict":"restricts publishing","source":"https://www.assemblyai.com/legal/terms-of-service","read":"2026-10-29","quote":"use or access the Services to develop a product or service that is competitive with any AssemblyAI product or service, or engage in competitive analysis or benchmarking;"}},{"id":"assemblyai-universal-streaming","company":"AssemblyAI","product":"Universal-Streaming (English / Multilingual)","layer":"stt","homepage":"https://www.assemblyai.com/","summary":"AssemblyAI's cheaper live speech-to-text model, English or six European languages.","open_weights":false,"facts":{"price":{"value":"$0.15 per hour (about $0.0025 per minute)","amount":0.0025,"currency":"USD","per":"audio_minute","tier":"pay as you go, streaming, per-minute figure derived from the hourly rate","source":"https://www.assemblyai.com/pricing","read":"2026-10-29","quote":"Universal-Streaming High-quality realtime model for English only transcription. Purpose-built for production voice applications with reliable accuracy, low latency, and cost-effectiveness. $0.15 /hr","note":"Billing docs (https://www.assemblyai.com/docs/billing-and-pricing): \"Streaming Speech-to-Text is billed on the total duration that your WebSocket connection stays open, not on the amount of audio you send.\" So idle open connections cost money. Multilingual is also $0.15/hr."},"latency":{"value":null,"source":"https://www.assemblyai.com/docs/streaming/universal-streaming","read":"2026-10-29","quote":"Fast","note":"Relative label only, no milliseconds."},"accuracy":{"value":null,"source":"https://www.assemblyai.com/docs/streaming/universal-streaming","read":"2026-10-29","quote":"Okay","note":"Relative entity-accuracy label only, no WER or dataset."},"en_gb":{"value":null,"source":"https://www.assemblyai.com/pricing","read":"2026-10-29","quote":"","note":"No mention of British English or accents on the pages read."},"streaming":{"value":"yes","source":"https://www.assemblyai.com/docs/streaming/universal-streaming","read":"2026-10-29","quote":"Universal-Streaming English","note":"Listed as a streaming model."},"languages":{"value":"English only (English model), English, Spanish, German, French, Portuguese, Italian (Multilingual)","source":"https://www.assemblyai.com/docs/streaming/universal-streaming","read":"2026-10-29","quote":"EN, ES, DE, FR, PT, IT","note":""},"data_region":{"value":"EU data zone endpoint (wss://streaming.eu.assemblyai.com/v3/ws) processes streaming audio and transcripts in the EU, the default endpoint routes to the nearest region, no UK region listed","source":"https://www.assemblyai.com/docs/streaming/endpoints-and-data-zones","read":"2026-10-29","quote":"Data Zone endpoints will process your audio and transcript data in the specified region.","note":"The security page snippet mentioned a Dublin data centre but the page text read did not say Dublin, so it is not recorded."},"compliance":{"value":"SOC 2 Type 2 report, signed BAA available (HIPAA), ISO 27001 and GDPR not confirmed on the pages read","source":"https://www.assemblyai.com/security","read":"2026-10-29","quote":"We maintain a SOC 2 Type 2 report covering security, availability, and confidentiality, audited annually by an independent firm.","note":"Security page also says: \"A signed BAA opts you out automatically.\" The pricing page does not state ISO 27001 or GDPR in its text, so those are not claimed here."}},"independent":[{"by":"Pipecat STT benchmark (semantic WER judged by Claude Sonnet 5.5)","metric":"semantic WER, mean across 1000 conversational voice-agent turns (lower is better)","value":"2.81%","source":"https://github.com/pipecat-ai/stt-benchmark","read":"2026-10-29","quote":"Columns: Vendor, Model, Transcripts, Perfect, WER Mean, Pooled WER, TTFS Median, TTFS P95, TTFS P99. Row: | AssemblyAI | universal-streaming-english | 99.8% | 66.8% | 2.81% | 2.42% | 256ms | 362ms | 417ms |","note":"Semantic WER counts only errors that change meaning for an LLM agent. Judge is an LLM (Claude), so scores depend on it. README says WER fell about 13% on average after the September 2026 judge change. Streaming services. The source calls the model \"AssemblyAI universal-streaming-english\"."},{"by":"Pipecat STT benchmark (semantic WER judged by Claude Sonnet 5.5)","metric":"time to final segment (TTFS), median","value":"256ms","source":"https://github.com/pipecat-ai/stt-benchmark","read":"2026-10-29","quote":"Columns: Vendor, Model, Transcripts, Perfect, WER Mean, Pooled WER, TTFS Median, TTFS P95, TTFS P99. Row: | AssemblyAI | universal-streaming-english | 99.8% | 66.8% | 2.81% | 2.42% | 256ms | 362ms | 417ms |","note":"TTFS measured from a shared end-of-speech anchor to the final transcript, run location and date not stated in README. The source calls the model \"AssemblyAI universal-streaming-english\"."}],"terms":{"verdict":"restricts publishing","source":"https://www.assemblyai.com/legal/terms-of-service","read":"2026-10-29","quote":"use or access the Services to develop a product or service that is competitive with any AssemblyAI product or service, or engage in competitive analysis or benchmarking;"}},{"id":"speechmatics-linden-1","company":"Speechmatics","product":"Linden 1 (Agent STT)","layer":"stt","homepage":"https://www.speechmatics.com/","summary":"Speechmatics' live speech-to-text model built for voice agents, sold as Agent STT.","open_weights":false,"facts":{"price":{"value":"$0.30 per hour (about $0.005 per minute)","amount":0.005,"currency":"USD","per":"audio_minute","tier":"pay as you go, real-time, per-minute figure derived from the hourly rate","source":"https://www.speechmatics.com/company/articles-and-news/meet-agent-stt-powered-by-linden-1-fast-accurate-speech-recognition","read":"2026-10-29","quote":"Launch pricing is $0.30 per hour, with further discounts available for high-volume usage.","note":"The pricing page data (https://www.speechmatics.com/pricing) also lists Linden 1 real-time at baseHourly 0.3, with tooltip \"Built for voice agents. Turn detection commits a final transcript the moment a speaker stops.\" The article quotes $0.16 after discounts, that is not a list price. Other real-time models on the pricing page data: Enhanced $0.80/hr, Standard $0.45/hr. A third-party rate tracker quoted different numbers, not used."},"latency":{"value":"End-of-speech to final transcript: under 350 ms mean and about 440 ms at P95 (vendor measurement)","source":"https://www.speechmatics.com/company/articles-and-news/meet-agent-stt-powered-by-linden-1-fast-accurate-speech-recognition","read":"2026-10-29","quote":"End-of-speech finalization is less than 350ms mean and ~440ms at P95 in our internal measurements.","note":"Measures time from the speaker stopping to the final transcript. The same article gives other figures (median ~350 ms, under 300 ms) and quotes Pipecat's benchmark at 369 ms median, which is a third-party figure that belongs in the independent section."},"accuracy":{"value":"1.05% pooled semantic word error rate (vendor-quoted)","source":"https://www.speechmatics.com/company/articles-and-news/meet-agent-stt-powered-by-linden-1-fast-accurate-speech-recognition","read":"2026-10-29","quote":"Median finalization is now ~350ms, at highest accuracy of 1.05% pooled semantic WER, for both the words and for who spoke them.","note":"Semantic WER counts only errors that change meaning. Speechmatics reports this on Pipecat's open benchmark (1,000 samples, smart-turn-data-v3.1), so it is a vendor-quoted figure, not independent. The Pipecat README as read on 29 Oct 2026 lists a pooled WER of 0.96% for linden-1, so the article's 1.05% is an earlier figure. Vendor also claims \"the highest accuracy we've measured across 55+ languages at that speed.\""},"en_gb":{"value":"yes, as an accent of its single Global English model","source":"https://docs.speechmatics.com/speech-to-text/languages","read":"2026-10-29","quote":"English spoken in the United Kingdom, United States, Australia, New Zealand, and by non-native speakers","note":"No en-GB code, one English model with spelling set by an output locale. Page: \"To standardize spelling, specify the output locale\"."},"streaming":{"value":"yes","source":"https://www.speechmatics.com/pricing","read":"2026-10-29","quote":"Built for voice agents.","note":"Listed in the real-time models list on the pricing page."},"languages":{"value":"55+ at launch, including bilingual and quadlingual models","source":"https://www.speechmatics.com/company/articles-and-news/meet-agent-stt-powered-by-linden-1-fast-accurate-speech-recognition","read":"2026-10-29","quote":"55+ at launch, including bilingual and quadlingual models","note":""},"data_region":{"value":"Region pinning available (US, Europe, Australia), a global endpoint is used otherwise. No UK region named.","source":"https://www.speechmatics.com/company/articles-and-news/meet-agent-stt-powered-by-linden-1-fast-accurate-speech-recognition","read":"2026-10-29","quote":"If you've got data residency requirements, you can pin to a region (US, Europe, Australia) instead.","note":"Security page says only \"Speechmatics have data processing environments located globally\" (https://www.speechmatics.com/security). On-premises containers are offered."},"compliance":{"value":"ISO/IEC 27001:2022, SOC 2 Type II, GDPR, HIPAA","source":"https://docs.speechmatics.com/administration/security-and-compliance","read":"2026-10-29","quote":"ISO/IEC 27001:2022","note":"Same page lists SOC 2 Type II, GDPR and HIPAA. Audit reports are behind a Trust Center NDA."}},"independent":[{"by":"Pipecat STT benchmark (semantic WER judged by Claude Sonnet 5.5)","metric":"semantic WER, mean across 1000 conversational voice-agent turns (lower is better)","value":"1.11%","source":"https://github.com/pipecat-ai/stt-benchmark","read":"2026-10-29","quote":"Columns: Vendor, Model, Transcripts, Perfect, WER Mean, Pooled WER, TTFS Median, TTFS P95, TTFS P99. Row: | Speechmatics | linden-1 | 99.5% | 83.9% | 1.11% | 0.96% | 369ms | 438ms | 690ms |","note":"Semantic WER counts only errors that change meaning for an LLM agent. Judge is an LLM (Claude), so scores depend on it. README says WER fell about 13% on average after the September 2026 judge change. Streaming services. The source calls the model \"Speechmatics linden-1\"."},{"by":"Pipecat STT benchmark (semantic WER judged by Claude Sonnet 5.5)","metric":"time to final segment (TTFS), median","value":"369ms","source":"https://github.com/pipecat-ai/stt-benchmark","read":"2026-10-29","quote":"Columns: Vendor, Model, Transcripts, Perfect, WER Mean, Pooled WER, TTFS Median, TTFS P95, TTFS P99. Row: | Speechmatics | linden-1 | 99.5% | 83.9% | 1.11% | 0.96% | 369ms | 438ms | 690ms |","note":"TTFS measured from a shared end-of-speech anchor to the final transcript, run location and date not stated in README. The source calls the model \"Speechmatics linden-1\"."}],"terms":{"verdict":"restricts other use","source":"https://www.speechmatics.com/legal/terms-of-service","read":"2026-10-29","quote":"Each party undertakes that it shall not at any time disclose to any person any Confidential Information except as permitted by section 6.2."}},{"id":"openai-gpt-live-transcribe","company":"OpenAI","product":"gpt-live-transcribe","layer":"stt","homepage":"https://platform.openai.com/","summary":"OpenAI's low-delay live speech-to-text model, used through its realtime transcription connection.","open_weights":false,"facts":{"price":{"value":"$0.017 per minute","amount":0.017,"currency":"USD","per":"audio_minute","tier":"pay as you go, live streaming","source":"https://developers.openai.com/api/docs/pricing","read":"2026-10-29","quote":"gpt-live-transcribe Live transcription - - $0.017 / minute","note":"Priced by audio duration, not tokens. The pricing table column is headed \"Estimated cost\"."},"latency":{"value":null,"source":"https://developers.openai.com/api/docs/models/gpt-live-transcribe","read":"2026-10-29","quote":"","note":"Page says it \"supports tunable latency\" but gives no milliseconds."},"accuracy":{"value":null,"source":"https://developers.openai.com/api/docs/models/gpt-live-transcribe","read":"2026-10-29","quote":"","note":"No word error rate or dataset given on the model page."},"en_gb":{"value":null,"source":"https://developers.openai.com/api/docs/models/gpt-live-transcribe","read":"2026-10-29","quote":"","note":"No English variant or accent handling stated on the model page, only language hints are mentioned."},"streaming":{"value":"yes","source":"https://developers.openai.com/api/docs/models/gpt-live-transcribe","read":"2026-10-29","quote":"Low-latency speech-to-text model for realtime transcription","note":""},"languages":{"value":null,"source":"https://developers.openai.com/api/docs/models/gpt-live-transcribe","read":"2026-10-29","quote":"","note":"The model page does not list supported languages."},"data_region":{"value":"Europe (EEA and Switzerland): regional storage and processing via eu.api.openai.com for audio transcription. UK: storage only, not processing. Both need approval for retention controls.","source":"https://developers.openai.com/api/docs/guides/your-data","read":"2026-10-29","quote":"GPT-Live supports data residency in the United States and Europe.","note":"The UK entry on the data residency table (gb.api.openai.com) shows storage Yes, processing No, and \"Requires MAM or ZDR\". The EU entry (eu.api.openai.com) shows storage and processing, also \"Requires MAM or ZDR\". Page: \"To use data residency with any region other than the United States, you must be approved for abuse monitoring controls, and execute a Modified Retention amendment.\""},"compliance":{"value":null,"source":"https://developers.openai.com/api/docs/guides/your-data","read":"2026-10-29","quote":"Zero Data Retention excludes customer content from abuse monitoring logs in the same way as Modified Abuse Monitoring.","note":"The your-data guide describes Zero Data Retention and Modified Abuse Monitoring controls. No SOC 2, ISO or HIPAA claim was confirmed: openai.com/security-and-privacy and /enterprise-privacy returned 403, so they were not read."}},"terms":{"verdict":"restricts other use","source":"https://web.archive.org/web/20261028230302/https://openai.com/policies/services-agreement/","read":"2026-10-29","quote":"You may use OpenAI models for research, software development, model evaluation and benchmarking, and product development, within these boundaries."}},{"id":"openai-gpt-realtime-whisper","company":"OpenAI","product":"gpt-realtime-whisper","layer":"stt","homepage":"https://platform.openai.com/","summary":"OpenAI's live speech-to-text model in the Whisper family, priced per audio minute.","open_weights":false,"facts":{"price":{"value":"$0.017 per minute","amount":0.017,"currency":"USD","per":"audio_minute","tier":"pay as you go, live streaming","source":"https://developers.openai.com/api/docs/pricing","read":"2026-10-29","quote":"gpt-realtime-whisper Live transcription - - $0.017 / minute","note":"Same price as gpt-live-transcribe. Priced by audio duration, not tokens. The pricing table column is headed \"Estimated cost\"."},"latency":{"value":null,"source":"https://developers.openai.com/api/docs/models/gpt-realtime-whisper","read":"2026-10-29","quote":"","note":"Page says developers can tune latency and accuracy, no milliseconds."},"accuracy":{"value":null,"source":"https://developers.openai.com/api/docs/models/gpt-realtime-whisper","read":"2026-10-29","quote":"","note":"No word error rate or dataset given on the model page."},"en_gb":{"value":null,"source":"https://developers.openai.com/api/docs/models/gpt-realtime-whisper","read":"2026-10-29","quote":"","note":"No English variant or accent handling stated on the model page, only language hints are mentioned."},"streaming":{"value":"yes","source":"https://developers.openai.com/api/docs/models/gpt-realtime-whisper","read":"2026-10-29","quote":"a streaming speech-to-text model for applications that need low-latency transcript deltas from live audio","note":""},"languages":{"value":null,"source":"https://developers.openai.com/api/docs/models/gpt-realtime-whisper","read":"2026-10-29","quote":"","note":"The model page does not list supported languages."},"data_region":{"value":"Europe (EEA and Switzerland): regional storage and processing via eu.api.openai.com for audio transcription. UK: storage only, not processing. Both need approval for retention controls.","source":"https://developers.openai.com/api/docs/guides/your-data","read":"2026-10-29","quote":"GPT-Live supports data residency in the United States and Europe.","note":"The UK entry on the data residency table (gb.api.openai.com) shows storage Yes, processing No, and \"Requires MAM or ZDR\". The EU entry (eu.api.openai.com) shows storage and processing, also \"Requires MAM or ZDR\". Page: \"To use data residency with any region other than the United States, you must be approved for abuse monitoring controls, and execute a Modified Retention amendment.\""},"compliance":{"value":null,"source":"https://developers.openai.com/api/docs/guides/your-data","read":"2026-10-29","quote":"Zero Data Retention excludes customer content from abuse monitoring logs in the same way as Modified Abuse Monitoring.","note":"The your-data guide describes Zero Data Retention and Modified Abuse Monitoring controls. No SOC 2, ISO or HIPAA claim was confirmed: openai.com/security-and-privacy and /enterprise-privacy returned 403, so they were not read."}},"independent":[{"by":"Pipecat STT benchmark (semantic WER judged by Claude Sonnet 5.5)","metric":"semantic WER, mean across 1000 conversational voice-agent turns (lower is better)","value":"2.41%","source":"https://github.com/pipecat-ai/stt-benchmark","read":"2026-10-29","quote":"Columns: Vendor, Model, Transcripts, Perfect, WER Mean, Pooled WER, TTFS Median, TTFS P95, TTFS P99. Row: | OpenAI | gpt-realtime-whisper | 100.0% | 73.3% | 2.41% | 2.30% | 740ms | 878ms | 1080ms |","note":"Semantic WER counts only errors that change meaning for an LLM agent. Judge is an LLM (Claude), so scores depend on it. README says WER fell about 13% on average after the September 2026 judge change. Streaming services. The source calls the model \"OpenAI gpt-realtime-whisper\"."},{"by":"Pipecat STT benchmark (semantic WER judged by Claude Sonnet 5.5)","metric":"time to final segment (TTFS), median","value":"740ms","source":"https://github.com/pipecat-ai/stt-benchmark","read":"2026-10-29","quote":"Columns: Vendor, Model, Transcripts, Perfect, WER Mean, Pooled WER, TTFS Median, TTFS P95, TTFS P99. Row: | OpenAI | gpt-realtime-whisper | 100.0% | 73.3% | 2.41% | 2.30% | 740ms | 878ms | 1080ms |","note":"TTFS measured from a shared end-of-speech anchor to the final transcript, run location and date not stated in README. The source calls the model \"OpenAI gpt-realtime-whisper\"."}],"terms":{"verdict":"restricts other use","source":"https://web.archive.org/web/20261028230302/https://openai.com/policies/services-agreement/","read":"2026-10-29","quote":"You may use OpenAI models for research, software development, model evaluation and benchmarking, and product development, within these boundaries."}},{"id":"openai-gpt-transcribe","company":"OpenAI","product":"gpt-transcribe","layer":"stt","homepage":"https://platform.openai.com/","summary":"OpenAI's newest accuracy-focused speech-to-text model for files and committed turns in a realtime session.","open_weights":false,"facts":{"price":{"value":"$0.0045 per minute","amount":0.0045,"currency":"USD","per":"audio_minute","tier":"pay as you go, file and Realtime turns","source":"https://developers.openai.com/api/docs/pricing","read":"2026-10-29","quote":"gpt-transcribe Transcription - - $0.0045 / minute","note":"Built for files and committed Realtime turns rather than word-by-word live output. The pricing table column is headed \"Estimated cost\"."},"latency":{"value":null,"source":"https://developers.openai.com/api/docs/models/gpt-transcribe","read":"2026-10-29","quote":"","note":"No latency figure on the model page."},"accuracy":{"value":null,"source":"https://developers.openai.com/api/docs/models/gpt-transcribe","read":"2026-10-29","quote":"","note":"No word error rate or dataset given on the model page."},"en_gb":{"value":null,"source":"https://developers.openai.com/api/docs/models/gpt-transcribe","read":"2026-10-29","quote":"","note":"No English variant or accent handling stated on the model page, only language hints are mentioned."},"streaming":{"value":"partly: streamed file transcripts and committed Realtime turns","source":"https://developers.openai.com/api/docs/models/gpt-transcribe","read":"2026-10-29","quote":"GPT Transcribe is a speech-to-text model for completed audio files, streamed file transcripts, and committed turns in Realtime sessions over WebSocket.","note":"Not clearly a live word-by-word model: the page describes files, streamed file transcripts and committed turns in Realtime sessions."},"languages":{"value":null,"source":"https://developers.openai.com/api/docs/models/gpt-transcribe","read":"2026-10-29","quote":"","note":"The model page does not list supported languages."},"data_region":{"value":"Europe (EEA and Switzerland): regional storage and processing via eu.api.openai.com for audio transcription. UK: storage only, not processing. Both need approval for retention controls.","source":"https://developers.openai.com/api/docs/guides/your-data","read":"2026-10-29","quote":"GPT-Live supports data residency in the United States and Europe.","note":"The UK entry on the data residency table (gb.api.openai.com) shows storage Yes, processing No, and \"Requires MAM or ZDR\". The EU entry (eu.api.openai.com) shows storage and processing, also \"Requires MAM or ZDR\". Page: \"To use data residency with any region other than the United States, you must be approved for abuse monitoring controls, and execute a Modified Retention amendment.\""},"compliance":{"value":null,"source":"https://developers.openai.com/api/docs/guides/your-data","read":"2026-10-29","quote":"Zero Data Retention excludes customer content from abuse monitoring logs in the same way as Modified Abuse Monitoring.","note":"The your-data guide describes Zero Data Retention and Modified Abuse Monitoring controls. No SOC 2, ISO or HIPAA claim was confirmed: openai.com/security-and-privacy and /enterprise-privacy returned 403, so they were not read."}},"terms":{"verdict":"restricts other use","source":"https://web.archive.org/web/20261028230302/https://openai.com/policies/services-agreement/","read":"2026-10-29","quote":"You may use OpenAI models for research, software development, model evaluation and benchmarking, and product development, within these boundaries."}},{"id":"openai-gpt-4o-transcribe","company":"OpenAI","product":"gpt-4o-transcribe","layer":"stt","homepage":"https://platform.openai.com/","summary":"OpenAI's earlier GPT-4o based speech-to-text model.","open_weights":false,"facts":{"price":{"value":"$0.006 per minute (estimated)","amount":0.006,"currency":"USD","per":"audio_minute","tier":"pay as you go, estimated per minute","source":"https://developers.openai.com/api/docs/pricing","read":"2026-10-29","quote":"gpt-4o-transcribe Transcription $2.50 $10.00 $0.006 / minute","note":"Token prices per 1M: $2.50 input, $10.00 output, $0.006 per minute is OpenAI's own estimate. gpt-4o-transcribe-diarize is the same price."},"latency":{"value":null,"source":"https://developers.openai.com/api/docs/pricing","read":"2026-10-29","quote":"","note":"No latency figure on the pricing page."},"accuracy":{"value":null,"source":"https://developers.openai.com/api/docs/pricing","read":"2026-10-29","quote":"","note":"No word error rate or dataset given on the model page."},"en_gb":{"value":null,"source":"https://developers.openai.com/api/docs/pricing","read":"2026-10-29","quote":"","note":"No English variant or accent handling stated on the model page, only language hints are mentioned."},"streaming":{"value":null,"source":"https://developers.openai.com/api/docs/pricing","read":"2026-10-29","quote":"","note":"Streaming support not checked for this model on a page We opened."},"languages":{"value":null,"source":"https://developers.openai.com/api/docs/pricing","read":"2026-10-29","quote":"","note":"The model page does not list supported languages."},"data_region":{"value":"Europe (EEA and Switzerland): regional storage and processing via eu.api.openai.com for audio transcription. UK: storage only, not processing. Both need approval for retention controls.","source":"https://developers.openai.com/api/docs/guides/your-data","read":"2026-10-29","quote":"GPT-Live supports data residency in the United States and Europe.","note":"The UK entry on the data residency table (gb.api.openai.com) shows storage Yes, processing No, and \"Requires MAM or ZDR\". The EU entry (eu.api.openai.com) shows storage and processing, also \"Requires MAM or ZDR\". Page: \"To use data residency with any region other than the United States, you must be approved for abuse monitoring controls, and execute a Modified Retention amendment.\""},"compliance":{"value":null,"source":"https://developers.openai.com/api/docs/guides/your-data","read":"2026-10-29","quote":"Zero Data Retention excludes customer content from abuse monitoring logs in the same way as Modified Abuse Monitoring.","note":"The your-data guide describes Zero Data Retention and Modified Abuse Monitoring controls. No SOC 2, ISO or HIPAA claim was confirmed: openai.com/security-and-privacy and /enterprise-privacy returned 403, so they were not read."}},"independent":[{"by":"Pipecat STT benchmark (semantic WER judged by Claude Sonnet 5.5)","metric":"semantic WER, mean across 1000 conversational voice-agent turns (lower is better)","value":"3.10%","source":"https://github.com/pipecat-ai/stt-benchmark","read":"2026-10-29","quote":"Columns: Vendor, Model, Transcripts, Perfect, WER Mean, Pooled WER, TTFS Median, TTFS P95, TTFS P99. Row: | OpenAI | gpt-4o-transcribe | 99.3% | 75.7% | 3.10% | 2.95% | 637ms | 965ms | 1655ms |","note":"Semantic WER counts only errors that change meaning for an LLM agent. Judge is an LLM (Claude), so scores depend on it. README says WER fell about 13% on average after the September 2026 judge change. Streaming services. The source calls the model \"OpenAI gpt-4o-transcribe\"."},{"by":"Pipecat STT benchmark (semantic WER judged by Claude Sonnet 5.5)","metric":"time to final segment (TTFS), median","value":"637ms","source":"https://github.com/pipecat-ai/stt-benchmark","read":"2026-10-29","quote":"Columns: Vendor, Model, Transcripts, Perfect, WER Mean, Pooled WER, TTFS Median, TTFS P95, TTFS P99. Row: | OpenAI | gpt-4o-transcribe | 99.3% | 75.7% | 3.10% | 2.95% | 637ms | 965ms | 1655ms |","note":"TTFS measured from a shared end-of-speech anchor to the final transcript, run location and date not stated in README. The source calls the model \"OpenAI gpt-4o-transcribe\"."}],"terms":{"verdict":"restricts other use","source":"https://web.archive.org/web/20261028230302/https://openai.com/policies/services-agreement/","read":"2026-10-29","quote":"You may use OpenAI models for research, software development, model evaluation and benchmarking, and product development, within these boundaries."}},{"id":"openai-gpt-4o-mini-transcribe","company":"OpenAI","product":"gpt-4o-mini-transcribe","layer":"stt","homepage":"https://platform.openai.com/","summary":"OpenAI's cheaper, smaller GPT-4o based speech-to-text model.","open_weights":false,"facts":{"price":{"value":"$0.003 per minute (estimated)","amount":0.003,"currency":"USD","per":"audio_minute","tier":"pay as you go, estimated per minute","source":"https://developers.openai.com/api/docs/pricing","read":"2026-10-29","quote":"gpt-4o-mini-transcribe Transcription $1.25 $5.00 $0.003 / minute","note":"Token prices per 1M: $1.25 input, $5.00 output, $0.003 per minute is OpenAI's own estimate."},"latency":{"value":null,"source":"https://developers.openai.com/api/docs/pricing","read":"2026-10-29","quote":"","note":"No latency figure on the pricing page."},"accuracy":{"value":null,"source":"https://developers.openai.com/api/docs/pricing","read":"2026-10-29","quote":"","note":"No word error rate or dataset given on the model page."},"en_gb":{"value":null,"source":"https://developers.openai.com/api/docs/pricing","read":"2026-10-29","quote":"","note":"No English variant or accent handling stated on the model page, only language hints are mentioned."},"streaming":{"value":null,"source":"https://developers.openai.com/api/docs/pricing","read":"2026-10-29","quote":"","note":"Streaming support not checked for this model on a page We opened."},"languages":{"value":null,"source":"https://developers.openai.com/api/docs/pricing","read":"2026-10-29","quote":"","note":"The model page does not list supported languages."},"data_region":{"value":"Europe (EEA and Switzerland): regional storage and processing via eu.api.openai.com for audio transcription. UK: storage only, not processing. Both need approval for retention controls.","source":"https://developers.openai.com/api/docs/guides/your-data","read":"2026-10-29","quote":"GPT-Live supports data residency in the United States and Europe.","note":"The UK entry on the data residency table (gb.api.openai.com) shows storage Yes, processing No, and \"Requires MAM or ZDR\". The EU entry (eu.api.openai.com) shows storage and processing, also \"Requires MAM or ZDR\". Page: \"To use data residency with any region other than the United States, you must be approved for abuse monitoring controls, and execute a Modified Retention amendment.\""},"compliance":{"value":null,"source":"https://developers.openai.com/api/docs/guides/your-data","read":"2026-10-29","quote":"Zero Data Retention excludes customer content from abuse monitoring logs in the same way as Modified Abuse Monitoring.","note":"The your-data guide describes Zero Data Retention and Modified Abuse Monitoring controls. No SOC 2, ISO or HIPAA claim was confirmed: openai.com/security-and-privacy and /enterprise-privacy returned 403, so they were not read."}},"terms":{"verdict":"restricts other use","source":"https://web.archive.org/web/20261028230302/https://openai.com/policies/services-agreement/","read":"2026-10-29","quote":"You may use OpenAI models for research, software development, model evaluation and benchmarking, and product development, within these boundaries."},"independent":[{"by":"mu-bench (arXiv 2609.32082)","metric":"English (en-US) utterance error rate on 8 kHz phone calls, batch, lower is better","value":"4.0%","source":"https://arxiv.org/abs/2609.32082","quote":"OpenAI GPT-4o Mini Transcribe | batch | en-US 4.0 | all 27.2 | p50 661 ms","note":"Table 5. Share of caller turns transcribed with an error. Batch mode, the one mode the paper ran for all four. The \"all\" column covers five languages (en-US, es-MX, tr-TR, vi-VN, zh-CN), where the spread is far wider.","read":"2026-10-30"}]},{"id":"elevenlabs-scribe-v2-realtime","company":"ElevenLabs","product":"Scribe v2 Realtime","layer":"stt","homepage":"https://elevenlabs.io/","summary":"ElevenLabs' live speech-to-text model, covering over 90 languages.","open_weights":false,"facts":{"price":{"value":"$0.39 per hour (about $0.0065 per minute)","amount":0.0065,"currency":"USD","per":"audio_minute","tier":"pay as you go, realtime, per-minute figure derived from the hourly rate","source":"https://elevenlabs.io/pricing/api","read":"2026-10-29","quote":"Scribe v2 Realtime Speech to Text $0.39 Price per hour","note":"The page prints hourly rates only. Add-ons: keyterm prompting $0.080/hr, transcript editing $0.120/hr. Batch Scribe v2 is $0.22/hr. The FAQ says speech-to-text is billed 'per audio minute' while tables are per hour, the page does not reconcile this."},"latency":{"value":"About 150 ms to first partial transcript (vendor figure)","source":"https://elevenlabs.io/docs/overview/models","read":"2026-10-29","quote":"Get partial transcriptions in ~150 milliseconds","note":"Time to first partial, vendor says it excludes application and network latency. The capabilities page shows \"Low latency (~150ms†)\" with an unexplained footnote."},"accuracy":{"value":"English falls in the vendor's 'Excellent (≤ 5% WER)' tier, no dataset named","source":"https://elevenlabs.io/docs/overview/capabilities/speech-to-text","read":"2026-10-29","quote":"Excellent (≤ 5% WER)","note":"ElevenLabs groups languages into WER tiers but does not name a benchmark dataset."},"en_gb":{"value":null,"source":"https://elevenlabs.io/docs/overview/capabilities/speech-to-text","read":"2026-10-29","quote":"","note":"Page lists English only, no en-GB or accent handling mentioned."},"streaming":{"value":"yes","source":"https://elevenlabs.io/docs/overview/capabilities/speech-to-text","read":"2026-10-29","quote":"Learn how to transcribe audio with ElevenLabs in realtime with WebSockets.","note":""},"languages":{"value":"90+ languages","source":"https://elevenlabs.io/docs/overview/models","read":"2026-10-29","quote":"delivers state-of-the-art accuracy in over 90 languages","note":""},"data_region":{"value":"Storage in the EU, India and Singapore as well as the US (default), Enterprise customers only. No UK region.","source":"https://elevenlabs.io/docs/overview/administration/data-residency","read":"2026-10-29","quote":"Data residency is an exclusive feature available to ElevenLabs' Enterprise customers.","note":""},"compliance":{"value":"HIPAA support when Zero Retention Mode and a BAA are in place, SOC 2 / GDPR not confirmed on a page We could open","source":"https://elevenlabs.io/enterprise","read":"2026-10-29","quote":"HIPAA support is available when Zero Retention Mode is enabled and a Business Associate Agreement (BAA) is in place.","note":"The docs compliance page returned 403 and the enterprise page does not mention SOC 2, GDPR or ISO 27001. Search snippets suggested SOC 2 and ISO 27001 claims but We did not open a page that states them."}},"independent":[{"by":"Pipecat STT benchmark (semantic WER judged by Claude Sonnet 5.5)","metric":"semantic WER, mean across 1000 conversational voice-agent turns (lower is better)","value":"3.05%","source":"https://github.com/pipecat-ai/stt-benchmark","read":"2026-10-29","quote":"Columns: Vendor, Model, Transcripts, Perfect, WER Mean, Pooled WER, TTFS Median, TTFS P95, TTFS P99. Row: | ElevenLabs | scribe_v2_realtime | 99.7% | 80.3% | 3.05% | 3.13% | 281ms | 348ms | 407ms |","note":"Semantic WER counts only errors that change meaning for an LLM agent. Judge is an LLM (Claude), so scores depend on it. README says WER fell about 13% on average after the September 2026 judge change. Streaming services. The source calls the model \"ElevenLabs scribe_v2_realtime\"."},{"by":"Pipecat STT benchmark (semantic WER judged by Claude Sonnet 5.5)","metric":"time to final segment (TTFS), median","value":"281ms","source":"https://github.com/pipecat-ai/stt-benchmark","read":"2026-10-29","quote":"Columns: Vendor, Model, Transcripts, Perfect, WER Mean, Pooled WER, TTFS Median, TTFS P95, TTFS P99. Row: | ElevenLabs | scribe_v2_realtime | 99.7% | 80.3% | 3.05% | 3.13% | 281ms | 348ms | 407ms |","note":"TTFS measured from a shared end-of-speech anchor to the final transcript, run location and date not stated in README. The source calls the model \"ElevenLabs scribe_v2_realtime\"."}],"terms":{"verdict":"restricts other use","source":"https://elevenlabs.io/terms-of-use","read":"2026-10-29","quote":"Using our Services or Output to research or develop products or services that compete with ours"}},{"id":"google-chirp-3","company":"Google Cloud","product":"Speech-to-Text V2 (Chirp 3)","layer":"stt","homepage":"https://cloud.google.com/speech-to-text","summary":"Google Cloud's speech-to-text service, Chirp 3 is its newest model, with live streaming through the V2 API.","open_weights":false,"facts":{"price":{"value":"$0.016 per minute for the first 500,000 minutes a month","amount":0.016,"currency":"USD","per":"audio_minute","tier":"pay as you go, V2 API, Standard recognition tier (first 500,000 minutes per month)","source":"https://cloud.google.com/speech-to-text/pricing","read":"2026-10-29","quote":"Standard 0 minute to 500,000 minute $0.016 / 1 minute, per 1 month / account","note":"Tiers fall to $0.01, $0.008 and $0.004 per minute at higher volumes. The page lists models in the Standard tier as: \"default, command_and_search, latest_short, latest_long, phone_call, video, chirp (Speech-to-Text V2 only)\". It says 'chirp', not 'Chirp 3', so applying the Standard rate to Chirp 3 is an inference. Billing is rounded up to 1-second increments. Dynamic batch is $0.003 per minute (not streaming)."},"latency":{"value":null,"source":"https://docs.cloud.google.com/speech-to-text/docs/models/chirp-3","read":"2026-10-29","quote":"offers the lowest latency","note":"Docs say the SUPERSHORT endpointing setting \"offers the lowest latency\" and the default \"introduces a small delay in the final response\", no milliseconds."},"accuracy":{"value":null,"source":"https://docs.cloud.google.com/speech-to-text/docs/transcription-model","read":"2026-10-29","quote":"provides enhanced accuracy and speed beyond earlier Chirp models","note":"No word error rate or dataset on the pages read."},"en_gb":{"value":"yes","source":"https://docs.cloud.google.com/speech-to-text/docs/languages","read":"2026-10-29","quote":"en-GB","note":"The language list shows en-GB (English, United Kingdom) with the chirp_3 model in the eu location. Quote is the locale code as shown in the table."},"streaming":{"value":"yes","source":"https://docs.cloud.google.com/speech-to-text/docs/models/chirp-3","read":"2026-10-29","quote":"Speech.StreamingRecognize (good for streaming and real-time audio)","note":"Listed as Supported for V2."},"languages":{"value":null,"source":"https://docs.cloud.google.com/speech-to-text/docs/languages","read":"2026-10-29","quote":"","note":"Language count not taken from a page We read."},"data_region":{"value":"eu and us multi-region locations for Chirp 3. London (europe-west2) not listed.","source":"https://docs.cloud.google.com/speech-to-text/docs/models/chirp-3","read":"2026-10-29","quote":"us (multi-region) GA eu (multi-region) GA","note":"Docs table lists \"us (multi-region)\" and \"eu (multi-region)\" as GA and says more regions are planned."},"compliance":{"value":null,"source":"https://cloud.google.com/speech-to-text/pricing","read":"2026-10-29","quote":"","note":"Compliance certifications were not checked on a page We opened."}},"terms":{"verdict":"restricts publishing","source":"https://cloud.google.com/terms/service-terms","read":"2026-10-29","quote":"Benchmarking. Customer may itself (but may not permit a third party to): (a) conduct benchmark tests of the Services (each a “Test”); and (b) publicly disclose the results of such Tests only if (i) the public disclosure includes all necessary information to replicate the Tests, and (ii) Customer allows Google to conduct benchmark tests of Customer's publicly available products or services and publicly disclose the results of such tests. Notwithstanding the foregoing, Customer may not do either of the following on behalf of a hyperscale public cloud provider without Google's prior written consent: (A) conduct (directly or through a third party) any Test or (B) disclose the results of any such Test."},"independent":[{"by":"mu-bench (arXiv 2609.32082)","metric":"English (en-US) utterance error rate on 8 kHz phone calls, batch, lower is better","value":"3.2%","source":"https://arxiv.org/abs/2609.32082","quote":"Google Chirp-3 | batch | en-US 3.2 | all 11.9 | p50 734 ms","note":"Table 5. Share of caller turns transcribed with an error. Batch mode, the one mode the paper ran for all four. The \"all\" column covers five languages (en-US, es-MX, tr-TR, vi-VN, zh-CN), where the spread is far wider.","read":"2026-10-30"}]},{"id":"azure-speech-realtime","company":"Microsoft Azure","product":"Azure Speech real-time speech to text","layer":"stt","homepage":"https://azure.microsoft.com/products/ai-services/ai-speech","summary":"Microsoft's standard live speech-to-text service, available in UK South.","open_weights":false,"facts":{"price":{"value":"$1.00 per hour (about $0.0167 per minute)","amount":0.0167,"currency":"USD","per":"audio_minute","tier":"pay as you go, S1 standard real-time, UK South","source":"https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20%27Foundry%20Tools%27%20and%20productName%20eq%20%27Azure%20Speech%27%20and%20armRegionName%20eq%20%27uksouth%27%20and%20contains(meterName,%27Speech%20To%20Text%27)","read":"2026-10-29","quote":"S1 Speech To Text | 1.0 1 Hour USD uksouth","note":"Read from Microsoft's public Azure retail prices API, because the pricing web page shows '$-' placeholders to scripts. Row: productName Azure Speech, skuName S1, meterName S1 Speech To Text, unit 1 Hour, USD, UK South, effective 2018-11-01. Per-minute figure derived. Fast transcription (files, not live) is $0.36/hr."},"latency":{"value":null,"source":"https://learn.microsoft.com/en-us/azure/ai-services/speech-service/speech-to-text","read":"2026-10-29","quote":"Real-time transcription: Instant transcription with intermediate results for live audio inputs.","note":"No latency figure published on the overview page."},"accuracy":{"value":null,"source":"https://learn.microsoft.com/en-us/azure/ai-services/speech-service/speech-to-text","read":"2026-10-29","quote":"","note":"No vendor word error rate or dataset on the overview page."},"en_gb":{"value":"yes","source":"https://learn.microsoft.com/en-us/azure/ai-services/speech-service/language-support?tabs=stt","read":"2026-10-29","quote":"en-GB English (United Kingdom)","note":"The locale table lists en-GB, its columns cover fast transcription and post-stream refinement, so real-time support for en-GB is implied by the page intro, not shown in its own column."},"streaming":{"value":"yes","source":"https://learn.microsoft.com/en-us/azure/ai-services/speech-service/speech-to-text","read":"2026-10-29","quote":"Real-time transcription: Instant transcription with intermediate results for live audio inputs.","note":""},"languages":{"value":null,"source":"https://learn.microsoft.com/en-us/azure/ai-services/speech-service/language-support?tabs=stt","read":"2026-10-29","quote":"","note":"The language-support page lists locales in a table but states no total, so no count is given."},"data_region":{"value":"Speech is stored and processed in the region where the resource is created, UK South and UK West are listed as Speech regions","source":"https://learn.microsoft.com/en-us/azure/ai-services/speech-service/regions","read":"2026-10-29","quote":"Azure Speech stores and processes speech data in the region where you create your resource.","note":""},"compliance":{"value":null,"source":"https://learn.microsoft.com/en-us/azure/ai-services/speech-service/regions","read":"2026-10-29","quote":"","note":"Compliance certifications not checked on a page We opened."}},"independent":[{"by":"Pipecat STT benchmark (semantic WER judged by Claude Sonnet 5.5)","metric":"semantic WER, mean across 1000 conversational voice-agent turns (lower is better)","value":"0.98%","source":"https://github.com/pipecat-ai/stt-benchmark","read":"2026-10-29","quote":"Columns: Vendor, Model, Transcripts, Perfect, WER Mean, Pooled WER, TTFS Median, TTFS P95, TTFS P99. Row: | Azure | N/A | 100.0% | 83.5% | 0.98% | 1.01% | 1016ms | 1345ms | 1791ms |","note":"Semantic WER counts only errors that change meaning for an LLM agent. Judge is an LLM (Claude), so scores depend on it. README says WER fell about 13% on average after the September 2026 judge change. Streaming services. The source calls the model \"Azure N/A\"."},{"by":"Pipecat STT benchmark (semantic WER judged by Claude Sonnet 5.5)","metric":"time to final segment (TTFS), median","value":"1016ms","source":"https://github.com/pipecat-ai/stt-benchmark","read":"2026-10-29","quote":"Columns: Vendor, Model, Transcripts, Perfect, WER Mean, Pooled WER, TTFS Median, TTFS P95, TTFS P99. Row: | Azure | N/A | 100.0% | 83.5% | 0.98% | 1.01% | 1016ms | 1345ms | 1791ms |","note":"TTFS measured from a shared end-of-speech anchor to the final transcript, run location and date not stated in README. The source calls the model \"Azure N/A\"."},{"by":"mu-bench (arXiv 2609.32082)","metric":"English (en-US) utterance error rate on 8 kHz phone calls, batch, lower is better","value":"2.8%","source":"https://arxiv.org/abs/2609.32082","quote":"Microsoft Azure Speech | batch | en-US 2.8 | all 17.0 | p50 315 ms","note":"Table 5. Share of caller turns transcribed with an error. Batch mode, the one mode the paper ran for all four. The \"all\" column covers five languages (en-US, es-MX, tr-TR, vi-VN, zh-CN), where the spread is far wider.","read":"2026-10-30"}],"terms":{"verdict":"restricts other use","source":"https://www.microsoft.com/licensing/terms/productoffering/MicrosoftAzure/MCA","read":"2026-10-29","quote":"If Customer offers a product or service competitive to an Online Service, by using the Online Service, Customer waives any restrictions on competitive use and benchmark testing in the terms governing its competitive products and services. If Customer offers a product or service competitive to an Online Service and discloses, directly or through third parties, any benchmarks or comparative tests or evaluations (each, a “Benchmark”) of any Online Service, Customer will, upon request from Microsoft, provide: (a) all information necessary to replicate such Benchmark; and (b) access to Customer's competitive products and services for Microsoft, directly or through third parties, to perform and disclose Benchmarks. If Customer does not intend to waive such restrictions in its terms of use and agree to these terms, Customer is not allowed to use the Online Service."}},{"id":"azure-mai-transcribe-2-streaming","company":"Microsoft Azure","product":"MAI-Transcribe-2-Streaming (preview)","layer":"stt","homepage":"https://ai.azure.com/catalog/models/MAI-Transcribe-2-Streaming","summary":"Microsoft's own new live speech-to-text model, in preview and not yet in UK regions.","open_weights":false,"facts":{"price":{"value":null,"amount":null,"currency":null,"per":"audio_minute","tier":"preview","source":"https://azure.microsoft.com/en-gb/pricing/details/cognitive-services/speech-services/","read":"2026-10-29","quote":"","note":"Azure pricing page lists MAI Transcribe-2 with a promotional offer through 12/31/2026 but every rate shows as '$-' to a script, and the Azure retail prices API returned no MAI meter. The model catalogue page says 'View pricing' with no figure."},"latency":{"value":null,"source":"https://ai.azure.com/catalog/models/MAI-Transcribe-2-Streaming","read":"2026-10-29","quote":"returns fast partial hypotheses followed by stable final results","note":"No numeric latency on the catalogue page."},"accuracy":{"value":null,"source":"https://ai.azure.com/catalog/models/MAI-Transcribe-2-Streaming","read":"2026-10-29","quote":"","note":"No word error rate or dataset on the catalogue page."},"en_gb":{"value":null,"source":"https://ai.azure.com/catalog/models/MAI-Transcribe-2-Streaming","read":"2026-10-29","quote":"","note":"Not stated."},"streaming":{"value":"yes","source":"https://ai.azure.com/catalog/models/MAI-Transcribe-2-Streaming","read":"2026-10-29","quote":"low-latency streaming speech recognition model that turns live audio into text as it is spoken","note":""},"languages":{"value":"60 languages","source":"https://ai.azure.com/catalog/models/MAI-Transcribe-2-Streaming","read":"2026-10-29","quote":"enabling responsive speech experiences across 60 languages","note":""},"data_region":{"value":"Not available in UK South, UK West or most EU regions, Sweden Central only with Global processing","source":"https://learn.microsoft.com/en-us/azure/ai-services/speech-service/regions","read":"2026-10-29","quote":"`mai-transcribe-2-streaming` is available only in the indicated regions.","note":"Voice Live table: uksouth and westeurope show '-' for mai-transcribe-2-streaming, swedencentral shows 'Global', US regions show 'Data zone'."},"compliance":{"value":null,"source":"https://learn.microsoft.com/en-us/azure/ai-services/speech-service/regions","read":"2026-10-29","quote":"","note":"Not checked."}},"terms":{"verdict":"restricts other use","source":"https://www.microsoft.com/licensing/terms/productoffering/MicrosoftAzure/MCA","read":"2026-10-29","quote":"If Customer offers a product or service competitive to an Online Service, by using the Online Service, Customer waives any restrictions on competitive use and benchmark testing in the terms governing its competitive products and services. If Customer offers a product or service competitive to an Online Service and discloses, directly or through third parties, any benchmarks or comparative tests or evaluations (each, a “Benchmark”) of any Online Service, Customer will, upon request from Microsoft, provide: (a) all information necessary to replicate such Benchmark; and (b) access to Customer's competitive products and services for Microsoft, directly or through third parties, to perform and disclose Benchmarks. If Customer does not intend to waive such restrictions in its terms of use and agree to these terms, Customer is not allowed to use the Online Service."}},{"id":"gladia-solaria","company":"Gladia","product":"Solaria (real-time)","layer":"stt","homepage":"https://www.gladia.io/","summary":"Gladia's speech-to-text API for live and recorded audio, hosted in the EU by default.","open_weights":false,"facts":{"price":{"value":"€0.68 per hour (€0.0113 per minute)","amount":0.0113,"currency":"EUR","per":"audio_minute","tier":"Self-serve, real-time","source":"https://www.gladia.io/pricing","read":"2026-10-29","quote":"0.68€/hr (0.0113€/min)","note":"Price is in euros. Pricing page names Solaria-3 and Solaria-1 but gives model-specific rates for async only, so which model applies to real-time is not stated. New accounts get €50 free credit."},"latency":{"value":"Under 103 ms on partial transcripts (vendor figure)","source":"https://www.gladia.io/solaria","read":"2026-10-29","quote":"The model delivers < 103 ms on partial transcription","note":"Measures partial (interim) transcripts, no method or conditions on the page. Page does not say which Solaria version."},"accuracy":{"value":null,"source":"https://www.gladia.io/solaria","read":"2026-10-29","quote":"Our latest benchmarks on Common Voice and FLEURS show top accuracy in EN, ES, FR and IT.","note":"Claim only, no number on the page (it links to a benchmarks page)."},"en_gb":{"value":null,"source":"https://www.gladia.io/solaria","read":"2026-10-29","quote":"Accurate across noise, accents, overlapping speakers & messy recordings.","note":"General accents claim, no en-GB model or setting stated."},"streaming":{"value":"yes","source":"https://www.gladia.io/pricing","read":"2026-10-29","quote":"0.68€/hr (0.0113€/min)","note":"Real-time price row."},"languages":{"value":"100 languages","source":"https://www.gladia.io/solaria","read":"2026-10-29","quote":"We support 100 languages, including 42 that are completely unsupported by alternative API vendors.","note":""},"data_region":{"value":"EU and US clusters, EU (France) is the default, region chosen in the API call. No UK region.","source":"https://www.gladia.io/blog/data-residency-for-voice-and-transcription-data-eu-us-and-ai-compliance","read":"2026-10-29","quote":"we use European infrastructure based in France by default to respect GDPR constraints.","note":""},"compliance":{"value":"SOC 2 Type II, ISO 27001, HIPAA, GDPR","source":"https://www.gladia.io/blog/data-residency-for-voice-and-transcription-data-eu-us-and-ai-compliance","read":"2026-10-29","quote":"We hold SOC 2 Type II and ISO 27001 certifications and comply with HIPAA and GDPR requirements.","note":""}},"terms":{"verdict":"restricts other use","source":"https://www.gladia.io/general-terms-of-use-gladia-sas","read":"2026-10-29","quote":"(i) use the Services or Output to develop competing products or services; (j) use any part of the Services or Output as input for external AI/ML tools;"}},{"id":"soniox-stt-rt-v5","company":"Soniox","product":"stt-rt-v5","layer":"stt","homepage":"https://soniox.com/","summary":"Soniox's live speech-to-text model for voice agents, priced by tokens but about $0.12 an hour.","open_weights":false,"facts":{"price":{"value":"About $0.12 per hour (about $0.002 per minute)","amount":0.002,"currency":"USD","per":"audio_minute","tier":"pay as you go, real-time (streaming), per-minute figure derived from the hourly rate","source":"https://soniox.com/pricing","read":"2026-10-29","quote":"Equivalent to about $0.10/hour for async (file) and $0.12/hour for real-time (streaming) transcription.","note":"Actually billed in tokens: $2.00 per 1M input audio tokens, $4.00 per 1M output text tokens, $4.00 per 1M input text tokens (custom context). Page says 1 hour of audio is about 30,000 input audio tokens and 1 hour of speech about 15,000 output text tokens. Page mentions no free credits."},"latency":{"value":null,"source":"https://soniox.com/docs/stt/models","read":"2026-10-29","quote":"Faster and more reliable semantic endpointing","note":"Release notes talk about faster endpoint detection, no milliseconds."},"accuracy":{"value":null,"source":"https://soniox.com/docs/stt/models","read":"2026-10-29","quote":"Higher real-time transcription accuracy across 60+ languages","note":"Claim only, no WER figure or dataset on the page."},"en_gb":{"value":null,"source":"https://soniox.com/docs/stt/concepts/supported-languages","read":"2026-10-29","quote":"","note":"Language table lists English (en) only, no dialect or en-GB mention."},"streaming":{"value":"yes","source":"https://soniox.com/docs/stt/models","read":"2026-10-29","quote":"stt-rt-v5 is the new Soniox real-time speech-to-text model for live audio.","note":""},"languages":{"value":"60+ languages","source":"https://soniox.com/docs/stt/concepts/supported-languages","read":"2026-10-29","quote":"supports transcription and translation in 60+ languages","note":""},"data_region":{"value":"EU region available on request, audio and transcripts stay in the EU. No UK region.","source":"https://soniox.com/europe","read":"2026-10-29","quote":"remain in the EU for both processing and storage","note":"EU access is \"enabled by request\" and the region is chosen when a project is created."},"compliance":{"value":"SOC 2 Type 2, ISO/IEC 27001:2022, GDPR, HIPAA","source":"https://soniox.com/docs/stt/security-and-privacy","read":"2026-10-29","quote":"Soniox meets industry-leading certification standards: SOC 2 Type 2","note":"The same compliance list on the security docs page also names ISO/IEC 27001:2022, GDPR and HIPAA. Docs also say audio and transcripts are not used to train Soniox models. Real-time audio is not stored unless you ask."}},"independent":[{"by":"Pipecat STT benchmark (semantic WER judged by Claude Sonnet 5.5)","metric":"semantic WER, mean across 1000 conversational voice-agent turns (lower is better)","value":"1.11%","source":"https://github.com/pipecat-ai/stt-benchmark","read":"2026-10-29","quote":"Columns: Vendor, Model, Transcripts, Perfect, WER Mean, Pooled WER, TTFS Median, TTFS P95, TTFS P99. Row: | Soniox | stt-rt-v5 | 99.8% | 83.6% | 1.11% | 1.09% | 260ms | 305ms | 313ms |","note":"Semantic WER counts only errors that change meaning for an LLM agent. Judge is an LLM (Claude), so scores depend on it. README says WER fell about 13% on average after the September 2026 judge change. Streaming services. The source calls the model \"Soniox stt-rt-v5\"."},{"by":"Pipecat STT benchmark (semantic WER judged by Claude Sonnet 5.5)","metric":"time to final segment (TTFS), median","value":"260ms","source":"https://github.com/pipecat-ai/stt-benchmark","read":"2026-10-29","quote":"Columns: Vendor, Model, Transcripts, Perfect, WER Mean, Pooled WER, TTFS Median, TTFS P95, TTFS P99. Row: | Soniox | stt-rt-v5 | 99.8% | 83.6% | 1.11% | 1.09% | 260ms | 305ms | 313ms |","note":"TTFS measured from a shared end-of-speech anchor to the final transcript, run location and date not stated in README. The source calls the model \"Soniox stt-rt-v5\"."}],"terms":{"verdict":"restricts publishing","source":"https://soniox.com/policies/terms-of-service","read":"2026-10-29","quote":"You may not publish benchmark results, performance comparisons, or evaluations of the Services in a false, misleading, deceptive, or commercially disparaging manner, or in a way that misrepresents the Services, testing methodology, results, or Soniox."}},{"id":"revai-reverb","company":"Rev.ai","product":"Reverb (streaming)","layer":"stt","homepage":"https://www.rev.ai/","summary":"Rev's speech-to-text API with a live streaming option, priced per hour of Reverb transcription.","open_weights":false,"facts":{"price":{"value":null,"amount":null,"currency":"USD","per":"audio_minute","tier":"streaming","source":"https://www.rev.ai/pricing","read":"2026-10-29","quote":"","note":"No separate streaming price published. The pricing page lists Reverb Transcription at $0.20 per hour (15-second minimum, one-second increments) and Reverb Foreign Language at $0.30 per hour, not labelled as streaming, it only links to the Streaming Speech to Text API docs. Docs pricing URL returned 404."},"latency":{"value":null,"source":"https://docs.rev.ai/api/streaming/","read":"2026-10-29","quote":"","note":"No latency figure on the streaming docs page read."},"accuracy":{"value":null,"source":"https://www.rev.ai/pricing","read":"2026-10-29","quote":"","note":"No accuracy claim on the pages read, the /reverb page returned 404."},"en_gb":{"value":null,"source":"https://docs.rev.ai/api/streaming/requests/","read":"2026-10-29","quote":"English en","note":"Streaming docs list English as 'en' only, defaulting to English, no en-GB. They name transcribers 'machine' and 'machine_v2', not Reverb."},"streaming":{"value":"yes","source":"https://docs.rev.ai/api/streaming/requests/","read":"2026-10-29","quote":"transcriber=machine_v2","note":"WebSocket streaming exists, whether it uses Reverb is not stated on the pages read."},"languages":{"value":null,"source":"https://docs.rev.ai/api/streaming/requests/","read":"2026-10-29","quote":"","note":"No full language list on the page read."},"data_region":{"value":null,"source":"https://www.rev.ai/security","read":"2026-10-29","quote":"","note":"Security page does not address data residency, it links to a 'Global Deployments' docs page We did not read."},"compliance":{"value":"SOC 2 Type II, SOC 3 report available, HIPAA* (enterprise tier), GDPR (DPA), CJIS","source":"https://www.rev.ai/security","read":"2026-10-29","quote":"We adhere to global data protection regulations","note":"Page text: \"including CJIS, HIPAA*, SOC 2 Type II compliance, and our SOC 3 report is available for review.\" Also: \"Rev's enterprise tier provides HIPAA-compliant* transcription services to protect sensitive healthcare data.\" The asterisk on HIPAA is not explained."}},"terms":{"verdict":"restricts publishing","source":"https://www.rev.com/legal/acceptable-use","read":"2026-10-29","quote":"Restrictions. Except as expressly permitted under this Agreement, Customer shall not itself, nor shall it permit any other party to: ... (iii) access the Platform for purposes of benchmarking or developing, marketing, selling or distributing any product or service that competes with or includes features substantially similar to the Platform;"}},{"id":"mistral-voxtral-mini-transcribe-realtime-3","company":"Mistral AI","product":"Voxtral Mini Transcribe Realtime 3","layer":"stt","homepage":"https://mistral.ai/","summary":"Mistral's hosted live speech-to-text model, with adjustable delay and turn detection.","open_weights":false,"facts":{"price":{"value":"$0.008 per minute","amount":0.008,"currency":"USD","per":"audio_minute","tier":"pay as you go, realtime","source":"https://docs.mistral.ai/inference/pricing","read":"2026-10-29","quote":"Voxtral Mini Transcribe Realtime 3 | $0.008 /Min | $0.0008 /Min | —","note":"Second figure is the cached-input rate. The previous Voxtral Mini Transcribe Realtime (v26.02, open weights) is $0.006 per minute through the API. Mistral's model overview marks the version 3 model 'Premier' (not open weights), the mistral.ai/pricing page names no Voxtral model. Quote is the model row of the pricing table (cells joined by |), the page renders the table, not a sentence."},"latency":{"value":"Configurable from 80 ms to 2,400 ms","source":"https://docs.mistral.ai/models/model-cards/voxtral-mini-transcribe-realtime-3-0","read":"2026-10-29","quote":"configurable latency (80 ms to 2,400 ms)","note":"A delay setting you choose, not a measured response time, shorter delay trades off accuracy."},"accuracy":{"value":null,"source":"https://docs.mistral.ai/models/model-cards/voxtral-mini-transcribe-realtime-3-0","read":"2026-10-29","quote":"","note":"No word error rate or dataset on the model card. For the earlier open v26.02 model Mistral says \"At 480ms delay, it stays within 1-2% word error rate\" of its batch model (see the open-weights row)."},"en_gb":{"value":null,"source":"https://docs.mistral.ai/capabilities/audio_transcription","read":"2026-10-29","quote":"","note":"Docs do not mention US or UK English variants."},"streaming":{"value":"yes","source":"https://docs.mistral.ai/models/overview","read":"2026-10-29","quote":"Realtime transcription model for agentic use-cases with configurable latency and turn detection.","note":"Quote is from the model overview page."},"languages":{"value":null,"source":"https://docs.mistral.ai/models/model-cards/voxtral-mini-transcribe-realtime-3-0","read":"2026-10-29","quote":"","note":"Not listed on the model card. The earlier realtime model covers 13 languages."},"data_region":{"value":null,"source":"https://docs.mistral.ai/inference/pricing","read":"2026-10-29","quote":"","note":"EU or UK hosting not checked on a Mistral page We opened."},"compliance":{"value":null,"source":"https://docs.mistral.ai/inference/pricing","read":"2026-10-29","quote":"","note":"Not checked on a Mistral page We opened."}},"terms":{"verdict":"no clause found","source":"https://legal.mistral.ai/terms/commercial-terms-of-service","read":"2026-10-29","quote":"\"Confidential Information\" means any non-public data, information, or material disclosed by either party (the \"Disclosing Party\") to the other (the \"Receiving Party\"), whether in written, oral, electronic, or other form, that is marked as confidential or, given the nature or circumstances of disclosure, should reasonably be understood as confidential."}},{"id":"mistral-voxtral-mini-transcribe-realtime-26-02","company":"Mistral AI","product":"Voxtral Mini Transcribe Realtime (v26.02, open weights)","layer":"stt","homepage":"https://huggingface.co/mistralai/Voxtral-Mini-3B-Realtime-2602","summary":"Mistral's open-weight live speech-to-text model you can host yourself, or use through Mistral's API.","open_weights":true,"facts":{"price":{"value":null,"amount":null,"currency":null,"per":"audio_minute","tier":"open weights, self-hosted; hosted API separate","source":"https://docs.mistral.ai/inference/pricing","read":"2026-10-29","quote":"","note":"Free to download, you pay for hosting. Mistral's hosted API charges $0.006 per minute for this version (row: Voxtral Mini Transcribe Realtime | $0.006 /Min)."},"latency":{"value":"Delay configurable down to under 200 ms","source":"https://mistral.ai/news/voxtral-transcribe-2","read":"2026-10-29","quote":"delay configurable down to sub-200ms","note":"A chosen delay setting for the streaming architecture, not a measured response time."},"accuracy":{"value":"Within 1 to 2% word error rate of the batch model at 480 ms delay, matches it at 2.4 seconds","source":"https://mistral.ai/news/voxtral-transcribe-2","read":"2026-10-29","quote":"At 480ms delay, it stays within 1-2% word error rate","note":"Relative to Mistral's own batch model, which it puts at about 4% WER on FLEURS. Vendor claim."},"en_gb":{"value":null,"source":"https://docs.mistral.ai/capabilities/audio_transcription","read":"2026-10-29","quote":"","note":"Docs do not mention English variants."},"streaming":{"value":"yes","source":"https://docs.mistral.ai/capabilities/audio_transcription","read":"2026-10-29","quote":"Voxtral Realtime is built for live applications.","note":""},"languages":{"value":"13 languages","source":"https://mistral.ai/news/voxtral-transcribe-2","read":"2026-10-29","quote":"natively multilingual, achieving strong transcription performance in 13 languages","note":""},"licence":{"value":"Apache 2.0","source":"https://docs.mistral.ai/models/model-cards/voxtral-mini-transcribe-realtime-26-02","read":"2026-10-29","quote":"Apache 2.0","note":""},"data_region":{"value":null,"source":"https://docs.mistral.ai/models/model-cards/voxtral-mini-transcribe-realtime-26-02","read":"2026-10-29","quote":"","note":"Open-weights model: the publisher states no data region. Where audio is processed depends on whoever hosts the model."},"compliance":{"value":null,"source":"https://docs.mistral.ai/models/model-cards/voxtral-mini-transcribe-realtime-26-02","read":"2026-10-29","quote":"","note":"Open-weights model: the publisher states no compliance claim for self-hosting. Certifications depend on whoever hosts the model."}},"independent":[{"by":"Pipecat STT benchmark (semantic WER judged by Claude Sonnet 5.5)","metric":"semantic WER, mean across 1000 conversational voice-agent turns (lower is better)","value":"4.49%","source":"https://github.com/pipecat-ai/stt-benchmark","read":"2026-10-29","quote":"Columns: Vendor, Model, Transcripts, Perfect, WER Mean, Pooled WER, TTFS Median, TTFS P95, TTFS P99. Row: | Mistral | voxtral-mini-transcribe-realtime-2602 | 99.3% | 67.7% | 4.49% | 5.11% | 525ms | 973ms | 1913ms |","note":"Semantic WER counts only errors that change meaning for an LLM agent. Judge is an LLM (Claude), so scores depend on it. README says WER fell about 13% on average after the September 2026 judge change. Streaming services. The source calls the model \"Mistral voxtral-mini-transcribe-realtime-2602\"."},{"by":"Pipecat STT benchmark (semantic WER judged by Claude Sonnet 5.5)","metric":"time to final segment (TTFS), median","value":"525ms","source":"https://github.com/pipecat-ai/stt-benchmark","read":"2026-10-29","quote":"Columns: Vendor, Model, Transcripts, Perfect, WER Mean, Pooled WER, TTFS Median, TTFS P95, TTFS P99. Row: | Mistral | voxtral-mini-transcribe-realtime-2602 | 99.3% | 67.7% | 4.49% | 5.11% | 525ms | 973ms | 1913ms |","note":"TTFS measured from a shared end-of-speech anchor to the final transcript, run location and date not stated in README. The source calls the model \"Mistral voxtral-mini-transcribe-realtime-2602\"."}]},{"id":"openai-whisper-large-v3","company":"OpenAI (open source)","product":"Whisper large-v3 / large-v3-turbo","layer":"stt","homepage":"https://github.com/openai/whisper","summary":"OpenAI's open speech-to-text model for recorded audio, it is not built for live streaming out of the box.","open_weights":true,"facts":{"price":{"value":null,"amount":null,"currency":null,"per":"audio_minute","tier":"open weights, self-hosted","source":"https://huggingface.co/openai/whisper-large-v3-turbo","read":"2026-10-29","quote":"","note":"Free to download, you pay for hosting. Hosted Whisper elsewhere is priced separately. The OpenAI pricing page read on 29 Oct 2026 has no Whisper row."},"latency":{"value":null,"source":"https://huggingface.co/openai/whisper-large-v3-turbo","read":"2026-10-29","quote":"","note":"No latency figure on the model card."},"accuracy":{"value":"Mean word error rate 7.83 on the Hugging Face Open ASR Leaderboard dataset collection (large-v3-turbo)","source":"https://huggingface.co/openai/whisper-large-v3-turbo","read":"2026-10-29","quote":"Mean Wer","note":"Figure sits in the model card's own evaluation metadata, so it is the card's claim, the independent leaderboard is another researcher's job."},"en_gb":{"value":null,"source":"https://huggingface.co/openai/whisper-large-v3-turbo","read":"2026-10-29","quote":"","note":"Model card does not mention British English or accents."},"streaming":{"value":"no (not out of the box)","source":"https://huggingface.co/openai/whisper-large-v3-turbo","read":"2026-10-29","quote":"Whisper models cannot be used for real-time transcription out of the box","note":"Others build near-real-time wrappers on top."},"languages":{"value":"99 languages","source":"https://huggingface.co/openai/whisper-large-v3-turbo","read":"2026-10-29","quote":"99 languages","note":""},"licence":{"value":"MIT","source":"https://huggingface.co/openai/whisper-large-v3-turbo","read":"2026-10-29","quote":"License: mit","note":"Card is for large-v3-turbo (809M parameters, decoder cut from 32 to 4 layers). large-v3 licence not separately read."},"data_region":{"value":null,"source":"https://huggingface.co/openai/whisper-large-v3-turbo","read":"2026-10-29","quote":"","note":"Open-weights model: the publisher states no data region. Where audio is processed depends on whoever hosts the model."},"compliance":{"value":null,"source":"https://huggingface.co/openai/whisper-large-v3-turbo","read":"2026-10-29","quote":"","note":"Open-weights model: the publisher states no compliance claim for self-hosting. Certifications depend on whoever hosts the model."}},"independent":[{"by":"Responsible ASR (Godambe et al., arXiv 2606.18659)","metric":"English WER on Indian-accented 8 kHz customer-support calls, overall","value":"28.7%","source":"https://arxiv.org/abs/2606.18659","read":"2026-10-29","quote":"Table 2 (English WER Agent / Customer / Overall; Hindi WER Agent / Customer / Overall). Row: Whisper-large-v3 1.5B 27.5 30.3 28.7 47.2 53.2 49.2","note":"Open-weights Whisper run by the authors, zero-shot. Customer channel 30.3%, agent channel 27.5%. Narrow-band 8 kHz telephony, spontaneous speech, proprietary recordings. Indian-accented English, not UK. The source calls the model \"Whisper-large-v3 (1.5B)\"."}]},{"id":"nvidia-parakeet-tdt-0-6b-v3","company":"NVIDIA","product":"Parakeet TDT 0.6B v3","layer":"stt","homepage":"https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3","summary":"NVIDIA's free 600-million-parameter speech-to-text model for 25 European languages, including English.","open_weights":true,"facts":{"price":{"value":null,"amount":null,"currency":null,"per":"audio_minute","tier":"open weights, self-hosted","source":"https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3","read":"2026-10-29","quote":"","note":"Free to download, you pay for hosting."},"latency":{"value":null,"source":"https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3","read":"2026-10-29","quote":"","note":"No latency figure on the model card. Its streaming example uses 2-second chunks with 2 seconds of right context."},"accuracy":{"value":"Average 6.34% word error rate on the Hugging Face Open ASR Leaderboard datasets (NVIDIA's own card)","source":"https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3","read":"2026-10-29","quote":"Huggingface Open-ASR-Leaderboard Model Avg WER AMI Earnings-22 GigaSpeech LS test-clean LS test-other SPGI Speech TEDLIUM-v3 VoxPopuli parakeet-tdt-0.6b-v3 6.34%","note":"Quote is the card's leaderboard table header and the average cell for this model. English on Fleurs 4.85%, LibriSpeech test-clean 1.93%. Self-reported by NVIDIA, clean read speech, not phone audio."},"en_gb":{"value":null,"source":"https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3","read":"2026-10-29","quote":"","note":"English only as 'English', no UK variant."},"streaming":{"value":"yes, through NVIDIA NeMo chunked inference","source":"https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3","read":"2026-10-29","quote":"To use parakeet models in streaming mode use this script","note":""},"languages":{"value":"25 European languages","source":"https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3","read":"2026-10-29","quote":"expanding language support from English to 25 European languages","note":""},"licence":{"value":"CC-BY-4.0","source":"https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3","read":"2026-10-29","quote":"GOVERNING TERMS: Use of this model is governed by the CC-BY-4.0 license.","note":""},"data_region":{"value":null,"source":"https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3","read":"2026-10-29","quote":"","note":"Open-weights model: the publisher states no data region. Where audio is processed depends on whoever hosts the model."},"compliance":{"value":null,"source":"https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3","read":"2026-10-29","quote":"","note":"Open-weights model: the publisher states no compliance claim for self-hosting. Certifications depend on whoever hosts the model."}}},{"id":"nvidia-canary-1b-v2","company":"NVIDIA","product":"Canary 1B v2","layer":"stt","homepage":"https://huggingface.co/nvidia/canary-1b-v2","summary":"NVIDIA's free speech-to-text and translation model for 25 European languages, aimed at recorded audio.","open_weights":true,"facts":{"price":{"value":null,"amount":null,"currency":null,"per":"audio_minute","tier":"open weights, self-hosted","source":"https://huggingface.co/nvidia/canary-1b-v2","read":"2026-10-29","quote":"","note":"Free to download, you pay for hosting."},"latency":{"value":null,"source":"https://huggingface.co/nvidia/canary-1b-v2","read":"2026-10-29","quote":"","note":"No latency figure on the model card."},"accuracy":{"value":"Open ASR Leaderboard mean 7.15% word error rate, Fleurs 8.40% across 25 languages (NVIDIA's own card)","source":"https://huggingface.co/nvidia/canary-1b-v2","read":"2026-10-29","quote":"Hugging Face Open ASR Leaderboard WER ↓ RTFx Mean AMI GigaSpeech LS Clean LS Other Earnings22 SPGISpech Tedlium Voxpopuli Canary-1b-v2 749 7.15 16.01 10.82 2.18 3.56 11.79 2.28 4.29 6.25","note":"Quote is the card's leaderboard table header and the row for this model (mean 7.15). Excludes punctuation and capitalisation errors. Self-reported by NVIDIA."},"en_gb":{"value":null,"source":"https://huggingface.co/nvidia/canary-1b-v2","read":"2026-10-29","quote":"","note":"No UK variant stated."},"streaming":{"value":null,"source":"https://huggingface.co/nvidia/canary-1b-v2","read":"2026-10-29","quote":"","note":"Card describes long-form inference with dynamic chunking and does not state streaming support."},"languages":{"value":"25 languages","source":"https://huggingface.co/nvidia/canary-1b-v2","read":"2026-10-29","quote":"built for high-quality speech transcription and translation across 25 European languages","note":""},"licence":{"value":"CC-BY-4.0","source":"https://huggingface.co/nvidia/canary-1b-v2","read":"2026-10-29","quote":"GOVERNING TERMS: Use of this model is governed by the CC-BY-4.0 license.","note":""},"data_region":{"value":null,"source":"https://huggingface.co/nvidia/canary-1b-v2","read":"2026-10-29","quote":"","note":"Open-weights model: the publisher states no data region. Where audio is processed depends on whoever hosts the model."},"compliance":{"value":null,"source":"https://huggingface.co/nvidia/canary-1b-v2","read":"2026-10-29","quote":"","note":"Open-weights model: the publisher states no compliance claim for self-hosting. Certifications depend on whoever hosts the model."}}},{"id":"kyutai-stt-2-6b-en","company":"Kyutai","product":"Kyutai STT 2.6B (English)","layer":"stt","homepage":"https://huggingface.co/kyutai/stt-2.6b-en","summary":"A free English-only streaming speech-to-text model from the French lab Kyutai.","open_weights":true,"facts":{"price":{"value":null,"amount":null,"currency":null,"per":"audio_minute","tier":"open weights, self-hosted","source":"https://huggingface.co/kyutai/stt-2.6b-en","read":"2026-10-29","quote":"","note":"Free to download, you pay for hosting."},"latency":{"value":"2.5 second delay","source":"https://huggingface.co/kyutai/stt-2.6b-en","read":"2026-10-29","quote":"2.5 second delay","note":"Fixed delay built into the English model per the card, not a measured end-to-end time."},"accuracy":{"value":"Mean word error rate 6.4 on the Hugging Face Open ASR Leaderboard datasets (card metadata)","source":"https://huggingface.co/kyutai/stt-2.6b-en","read":"2026-10-29","quote":"Mean Wer View evaluation results source 6.4","note":"Quote is the card's evaluation metadata block (mean WER 6.4). AMI WER 12.17. Self-reported on the card."},"en_gb":{"value":null,"source":"https://huggingface.co/kyutai/stt-2.6b-en","read":"2026-10-29","quote":"","note":"English only, no UK variant stated."},"streaming":{"value":"yes","source":"https://huggingface.co/kyutai/stt-2.6b-en","read":"2026-10-29","quote":"This is a model for streaming speech-to-text (STT, also known as automatic speech recognition, ASR).","note":""},"languages":{"value":"English only","source":"https://huggingface.co/kyutai/stt-2.6b-en","read":"2026-10-29","quote":"English for kyutai/stt-2.6b-en","note":""},"licence":{"value":"CC-BY 4.0 (model weights)","source":"https://huggingface.co/kyutai/stt-2.6b-en","read":"2026-10-29","quote":"Model weights are licensed under CC-BY 4.0","note":""},"data_region":{"value":null,"source":"https://huggingface.co/kyutai/stt-2.6b-en","read":"2026-10-29","quote":"","note":"Open-weights model: the publisher states no data region. Where audio is processed depends on whoever hosts the model."},"compliance":{"value":null,"source":"https://huggingface.co/kyutai/stt-2.6b-en","read":"2026-10-29","quote":"","note":"Open-weights model: the publisher states no compliance claim for self-hosting. Certifications depend on whoever hosts the model."}}},{"id":"elevenlabs-flash-v2-5","company":"ElevenLabs","product":"Flash v2.5","layer":"tts","homepage":"https://elevenlabs.io/","summary":"ElevenLabs' fastest speech model, built for live conversation.","open_weights":false,"facts":{"price":{"value":"$0.04 per 1,000 characters","amount":0.04,"currency":"USD","per":"1k_characters","tier":"pay as you go, Flash / Turbo","source":"https://elevenlabs.io/pricing/api","read":"2026-10-29","quote":"Flash / Turbo Text to Speech $0.04 Price per 1K characters","note":"FAQ on the same page: API usage is billed in US dollars, not credits. Plans (Starter $6, Creator $22, Pro $99 a month) bundle characters at the same per-character rate."},"latency":{"value":"~75ms","source":"https://elevenlabs.io/docs/overview/models","read":"2026-10-29","quote":"Ultra-low latency (~75ms†)","note":"Vendor figure. Footnote on the same page: \"† Excluding application & network latency.\" No method given beyond that."},"en_gb_voices":{"value":null,"source":"https://elevenlabs.io/pricing/api","read":"2026-10-29","quote":"","note":"Not counted. No page checked gives a count of British English voices, the voice library is not a static page."},"languages":{"value":"32","source":"https://elevenlabs.io/docs/overview/models","read":"2026-10-29","quote":"Ultra-low latency (~75ms†) 32 languages supported 40,000 character limit","note":"Flash v2.5 entry on the models page (latency, languages, character limit)."},"data_region":{"value":"US by default, EU, India and Singapore storage on Enterprise only","source":"https://elevenlabs.io/docs/overview/administration/data-residency","read":"2026-10-29","quote":"Data residency is an exclusive feature available to ElevenLabs' Enterprise customers.","note":"Page: customer data is stored in the U.S. by default."},"compliance":{"value":"SOC 2 Type II, ISO 27001, HIPAA, PCI DSS L1, GDPR, zero retention mode","source":"https://elevenlabs.io/agents/enterprise-conversational-ai","read":"2026-10-29","quote":"SOC 2 Type II, ISO 27001, HIPAA, PCI DSS L1, and GDPR certified, with zero retention mode, VPC deployment, and end-to-end encryption.","note":"This sentence is on the ElevenAgents (voice agent) enterprise page, not a TTS-only page. The trust centre (compliance.elevenlabs.io) loaded without any certification list. Check scope before quoting for TTS alone."}},"terms":{"verdict":"restricts other use","source":"https://elevenlabs.io/terms-of-use","read":"2026-10-29","quote":"Using our Services or Output to research or develop products or services that compete with ours"}},{"id":"elevenlabs-v4-turbo","company":"ElevenLabs","product":"Eleven v4 Turbo","layer":"tts","homepage":"https://elevenlabs.io/","summary":"ElevenLabs' newest fast speech model, aimed at expressive live agents.","open_weights":false,"facts":{"price":{"value":"$0.04 per 1,000 characters","amount":0.04,"currency":"USD","per":"1k_characters","tier":"pay as you go, v4 Turbo","source":"https://elevenlabs.io/pricing/api","read":"2026-10-29","quote":"v4 Turbo Text to Speech New $0.04 Price per 1K characters Get started Our best model for real-time and agents Ultra-low latency (~100ms)","note":"Eleven v4 (the flagship, no latency figure) is $0.08 per 1K characters on the same page."},"latency":{"value":"~100ms median inference","source":"https://elevenlabs.io/docs/overview/models","read":"2026-10-29","quote":"Ultra-low latency (median inference latency of ~100ms†)","note":"Vendor figure, median inference time. Footnote: \"† Excluding application & network latency.\""},"en_gb_voices":{"value":null,"source":"https://elevenlabs.io/pricing/api","read":"2026-10-29","quote":"","note":"Not counted. No page checked gives a count of British English voices, the voice library is not a static page."},"languages":{"value":"90+","source":"https://elevenlabs.io/docs/overview/models","read":"2026-10-29","quote":"Ultra-low latency (median inference latency of ~100ms†) Exceptional voice cloning capabilities 90+ languages supported","note":"Eleven v4 Turbo entry on the models page."},"data_region":{"value":"US by default, EU, India and Singapore storage on Enterprise only","source":"https://elevenlabs.io/docs/overview/administration/data-residency","read":"2026-10-29","quote":"Data residency is an exclusive feature available to ElevenLabs' Enterprise customers.","note":"Page: customer data is stored in the U.S. by default."},"compliance":{"value":"SOC 2 Type II, ISO 27001, HIPAA, PCI DSS L1, GDPR, zero retention mode","source":"https://elevenlabs.io/agents/enterprise-conversational-ai","read":"2026-10-29","quote":"SOC 2 Type II, ISO 27001, HIPAA, PCI DSS L1, and GDPR certified, with zero retention mode, VPC deployment, and end-to-end encryption.","note":"This sentence is on the ElevenAgents (voice agent) enterprise page, not a TTS-only page. The trust centre (compliance.elevenlabs.io) loaded without any certification list. Check scope before quoting for TTS alone."}},"terms":{"verdict":"restricts other use","source":"https://elevenlabs.io/terms-of-use","read":"2026-10-29","quote":"Using our Services or Output to research or develop products or services that compete with ours"}},{"id":"cartesia-sonic-3-6","company":"Cartesia","product":"Sonic 3.6","layer":"tts","homepage":"https://cartesia.ai/","summary":"Cartesia's current speech model for live voice agents.","open_weights":false,"facts":{"price":{"value":"Credit plans only: Pro is $5 a month for 100K credits, about $0.05 per 1,000 characters","amount":0.05,"currency":"USD","per":"1k_characters","tier":"Pro plan, effective rate (100K credits for $5)","source":"https://cartesia.ai/pricing","read":"2026-10-29","quote":"1 credit equals 1 character.","note":"No pay-as-you-go option on the pricing page. Cheapest paid plan is Pro, $5 for 100K credits = $50 per 1M. Startup $49 for 1.25M ($39.20 per 1M), Scale $299 for 8M ($37.38 per 1M). Overage when enabled: Pro $65, Startup $45, Scale $38 per 1M credits. Free plan 20K credits. Page also says one minute of audio needs 750-800 credits."},"latency":{"value":"first audio in under 90ms","source":"https://cartesia.ai/product/text-to-speech-tts","read":"2026-10-29","quote":"first audio in under 90ms","note":"Vendor claim on the product page for Sonic, with no method or conditions given. The Sonic 3.6 docs page and the pricing page give no millisecond figure."},"en_gb_voices":{"value":"yes, at least 2 named (Gemma, Archie), more in the voice library","source":"https://docs.cartesia.ai/build-with-cartesia/tts-models/latest","read":"2026-10-29","quote":"Gemma - en-GB Female (ID: 62ae83ad-4f6a-430b-af41-a9bede9286ca)","note":"Docs recommend two en-GB voices (Gemma female, Archie male) and point to the Voice Library for more, library total not counted."},"languages":{"value":"44","source":"https://docs.cartesia.ai/build-with-cartesia/tts-models/latest","read":"2026-10-29","quote":"Sonic 3.6 (sonic-3.6) is our fastest, most natural text-to-speech model, with native support for 44 languages.","note":""},"data_region":{"value":null,"source":"https://cartesia.ai/pricing","read":"2026-10-29","quote":"","note":"Pricing and model docs pages state no data region. On-premise or VPC deployment is mentioned for sales contact only."},"compliance":{"value":"SOC 2 Type II, DPAs and BAAs on Enterprise","source":"https://cartesia.ai/pricing","read":"2026-10-29","quote":"Yes, Cartesia is SOC 2 Type II certified.","note":"Enterprise tier lists DPAs and BAAs for compliance."}},"terms":{"verdict":"restricts other use","source":"https://www.cartesia.ai/legal/terms","read":"2026-10-29","quote":"Use any part of the Services or any Output to research and develop products, models and services that compete with Cartesia;"}},{"id":"deepgram-aura-2","company":"Deepgram","product":"Aura-2","layer":"tts","homepage":"https://deepgram.com/","summary":"Deepgram's streaming text-to-speech for business voice agents.","open_weights":false,"facts":{"price":{"value":"$0.030 per 1,000 characters","amount":0.03,"currency":"USD","per":"1k_characters","tier":"pay as you go","source":"https://deepgram.com/pricing","read":"2026-10-29","quote":"$0.030 /1k characters","note":"Pay as you go. Growth plan is $0.027 per 1,000 characters."},"latency":{"value":null,"source":"https://developers.deepgram.com/docs/text-to-speech","read":"2026-10-29","quote":"","note":"No time-to-first-audio figure on the pricing or TTS docs pages. Docs say: the audio is streamed back so playback can begin when the first byte arrives."},"en_gb_voices":{"value":"2 (Draco male, Pandora female)","source":"https://developers.deepgram.com/docs/tts-models","read":"2026-10-29","quote":"aura-2-pandora-en pandora feminine Adult en-gb British","note":"Two en-gb rows in the Aura-2 English voice table: aura-2-draco-en and aura-2-pandora-en."},"languages":{"value":"English, Spanish, German, French, Dutch, Italian, Japanese","source":"https://developers.deepgram.com/docs/tts-models","read":"2026-10-29","quote":"English (en) - American, British, Australian, Irish, Filipino accents Spanish (es) - Mexican, Peninsular, Colombian, Latin American accents German (de) French (fr) Dutch (nl) Italian (it) Japanese (ja)","note":""},"data_region":{"value":"EU endpoint available (api.eu.deepgram.com)","source":"https://deepgram.com/pricing","read":"2026-10-29","quote":"We offer a dedicated EU endpoint to ensure data processing stays within the European Union (api.eu.deepgram.com).","note":"No UK-specific region stated."},"compliance":{"value":"SOC 2 Type 1 and 2, HIPAA (BAA on Enterprise), GDPR ready, CCPA, PCI","source":"https://deepgram.com/pricing","read":"2026-10-29","quote":"SOC 2 Type 1 & Type 2 Certified Independently audited security controls and procedures to ensure your data is protected in the cloud.","note":""}},"terms":{"verdict":"restricts publishing","source":"https://deepgram.com/terms","read":"2026-10-29","quote":"use or access our Services or any Output for competitive purposes, including model training, benchmarking and other competitive analysis, or developing competing models, products or services;"}},{"id":"openai-gpt-4o-mini-tts","company":"OpenAI","product":"gpt-4o-mini-tts","layer":"tts","homepage":"https://openai.com/api/","summary":"OpenAI's text-to-speech model, priced by tokens. Scheduled to shut down on 6 January 2027.","open_weights":false,"facts":{"price":{"value":"$12.00 per 1M audio output tokens (plus $0.60 per 1M text input tokens)","amount":12,"currency":"USD","per":"1m_audio_tokens_output","tier":"pay as you go","source":"https://developers.openai.com/api/docs/models/gpt-4o-mini-tts","read":"2026-10-29","quote":"Audio tokens Per 1M tokens Output $12.00","note":"Priced per token, no per-minute or per-character figure printed on the page. Older tts-1 is $15.00 and tts-1-hd $30.00 per 1M characters. Docs page says all three TTS models will shut down on January 6, 2027 and recommends gpt-realtime-2.1-mini through the Realtime API (priced at $10.00 per 1M audio input and $20.00 per 1M audio output tokens). OpenAI's new voice sessions model gpt-live-1 is listed at $0.05 per minute."},"price_text_in":{"value":"$0.60 per 1M text input tokens","amount":0.6,"currency":"USD","per":"1m_tokens_input","tier":"pay as you go","source":"https://developers.openai.com/api/docs/models/gpt-4o-mini-tts","read":"2026-10-29","quote":"Text tokens Per 1M tokens Input $0.60","note":"Same model page as the audio output price."},"latency":{"value":null,"source":"https://developers.openai.com/api/docs/guides/text-to-speech","read":"2026-10-29","quote":"","note":"Docs give no figure for gpt-4o-mini-tts. Only a relative claim that tts-1 is lower latency than tts-1-hd, and advice to use WAV or PCM for the fastest response."},"en_gb_voices":{"value":"no dedicated British voices, accent is steered by prompt","source":"https://developers.openai.com/api/docs/guides/text-to-speech","read":"2026-10-29","quote":"gpt-4o-mini-tts lets you use prompts to control these aspects of speech: Accent Emotional range Intonation Impressions Speed of speech Tone Whispering","note":"The voice list on the page names 13 voices (alloy, ash, ballad, coral, echo, fable, nova, onyx, sage, shimmer, verse, marin, cedar), although the page intro says 11 built-in voices. The page does not mention British English by name."},"languages":{"value":"57 languages listed (follows Whisper language support)","source":"https://developers.openai.com/api/docs/guides/text-to-speech","read":"2026-10-29","quote":"The TTS model generally follows the Whisper model in terms of language support.","note":"The page lists the Whisper languages and says voices are optimized for English."},"data_region":{"value":"UK: regional storage yes, regional processing no. Europe (EEA + Switzerland) is listed per endpoint.","source":"https://developers.openai.com/api/docs/guides/your-data","read":"2026-10-29","quote":"United Kingdom gb.api.openai.com Storage Yes Processing No Requires MAM or ZDR","note":"Europe (EEA + Switzerland) regional processing requires Zero Data Retention or similar approved controls. The audio speech endpoint is listed as ZDR-eligible. Non-US regions need approval and a Modified Retention amendment."},"compliance":{"value":null,"source":"https://developers.openai.com/api/docs/guides/your-data","read":"2026-10-29","quote":"","note":"Trust centre pages returned 403. Zero Data Retention is available by approval and covers the Audio speech endpoint."}},"terms":{"verdict":"restricts other use","source":"https://web.archive.org/web/20261028230302/https://openai.com/policies/services-agreement/","read":"2026-10-29","quote":"You may use OpenAI models for research, software development, model evaluation and benchmarking, and product development, within these boundaries."}},{"id":"rime-mist-v3","company":"Rime","product":"Mist v3","layer":"tts","homepage":"https://www.rime.ai/","summary":"Rime's low-latency speech model for live calls.","open_weights":false,"facts":{"price":{"value":"$0.030 per minute","amount":0.03,"currency":"USD","per":"audio_minute","tier":"Flex, pay as you go","source":"https://www.rime.ai/pricing","read":"2026-10-29","quote":"Mist v3 Text to speech $0.030 / min","note":"Flex plan, pay as you go. Scale plan is 20% off ($0.024). FAQ: TTS is metered by characters generated, about 1K characters per minute."},"latency":{"value":"37 ms median (P50), 56 ms P90 time to first audio","source":"https://www.rime.ai/pricing","read":"2026-10-29","quote":"TTFA (P50) Time to First Audio at 1 concurrency Coda v2 Sub-100 ms Mist v3 37 ms","note":"Vendor benchmark at 1 concurrent request, docs say 'approximately 37ms P50 in Rime's benchmark'. Method not described further."},"en_gb_voices":{"value":"none described as British","source":"https://docs.rime.ai/docs/voices-mist-v3.md","read":"2026-10-29","quote":"The public catalog publishes 78 Mist v3 voices.","note":"The voice descriptions on the Mist v3 voices page mention no British or UK accent."},"languages":{"value":"English, French, German, Spanish","source":"https://docs.rime.ai/api-reference/models","read":"2026-10-29","quote":"Mist v3 serves English, French, German, and Spanish.","note":""},"data_region":{"value":null,"source":"https://docs.rime.ai/api-reference/models","read":"2026-10-29","quote":"route each request to the nearest Rime region","note":"No region list and no UK or EU statement. Enterprise offers cloud, VPC or on-prem."},"compliance":{"value":"HIPAA BAA and SOC 2 Type II on Enterprise only","source":"https://www.rime.ai/pricing","read":"2026-10-29","quote":"BAA (HIPAA) and SOC 2 Type II report","note":"Page lists these on the Enterprise plan, not Free, Flex or Scale."}},"terms":{"verdict":"restricts other use","source":"https://www.rime.ai/terms","read":"2026-10-29","quote":"use Output to train any artificial intelligence or machine learning model;"}},{"id":"rime-coda-v2","company":"Rime","product":"Coda v2","layer":"tts","homepage":"https://www.rime.ai/","summary":"Rime's newer, more natural speech model that Rime recommends for new apps.","open_weights":false,"facts":{"price":{"value":"$0.050 per minute","amount":0.05,"currency":"USD","per":"audio_minute","tier":"Flex, pay as you go","source":"https://www.rime.ai/pricing","read":"2026-10-29","quote":"Coda v2 Text to speech $0.050 / min","note":"Flex plan, pay as you go. Scale plan is 20% off. Earlier Rime model Arcana does not appear on the pricing or models pages any more."},"latency":{"value":"sub-100 ms median (P50) time to first audio at 1 concurrent request (vendor benchmark)","source":"https://www.rime.ai/pricing","read":"2026-10-29","quote":"TTFA (P50) Time to First Audio at 1 concurrency Coda v2 Sub-100 ms Mist v3 37 ms","note":"Vendor benchmark, method not described. Separately, Rime docs say sub-100ms model latency applies to the GPU engine (self-hosted or on-prem) and cloud API users typically add 25 to 50ms of network round-trip (docs.rime.ai/api-reference/models)."},"en_gb_voices":{"value":"11 labelled GB","source":"https://docs.rime.ai/docs/voices-coda.md","read":"2026-10-29","quote":"A calm, even-keeled younger British voice, quick on its feet.","note":"Catalogue lists 11 voices with accent GB (albion, bogg, fuzz, frost, greyline, meadow, northvale, rainy, saturn, vashti, wren)."},"languages":{"value":"10","source":"https://docs.rime.ai/api-reference/models","read":"2026-10-29","quote":"Ten cloud API languages: English, Arabic, Dutch, French, German, Hindi, Italian, Japanese, Portuguese, and Spanish.","note":""},"data_region":{"value":null,"source":"https://docs.rime.ai/api-reference/models","read":"2026-10-29","quote":"route each request to the nearest Rime region","note":"No region list and no UK or EU statement. Enterprise offers cloud, VPC or on-prem."},"compliance":{"value":"HIPAA BAA and SOC 2 Type II on Enterprise only","source":"https://www.rime.ai/pricing","read":"2026-10-29","quote":"BAA (HIPAA) and SOC 2 Type II report","note":"Page lists these on the Enterprise plan, not Free, Flex or Scale."}},"terms":{"verdict":"restricts other use","source":"https://www.rime.ai/terms","read":"2026-10-29","quote":"use Output to train any artificial intelligence or machine learning model;"}},{"id":"hume-octave-2","company":"Hume AI","product":"Octave 2 (preview)","layer":"tts","homepage":"https://www.hume.ai/","summary":"Hume's expressive speech model. Hume is closing its text-to-speech and voice APIs on 13 November 2026, so it cannot be used after that date.","open_weights":false,"facts":{"price":{"value":null,"source":"https://www.hume.ai/pricing","read":"2026-10-29","quote":"","note":"No Octave TTS price on the pages that loaded: hume.ai/pricing shows no rates and the docs page says rate limits depend on the subscription tier. Not stated whether any price is current given the shutdown."},"latency":{"value":"~100ms (Octave 2), ~200ms (Octave 1), not including network","source":"https://dev.hume.ai/docs/text-to-speech-tts/overview","read":"2026-10-29","quote":"Octave 2 (preview) generates high-quality audio with latencies as low as ~100ms (not including network transit)","note":"Vendor statement of model latency excluding network transit."},"shutdown":{"value":"Hume is sunsetting the TTS and EVI APIs, access ends 13 November 2026","source":"https://dev.hume.ai/docs/text-to-speech-tts/overview","read":"2026-10-29","quote":"Access ends November 13, 2026 at 12:01 a.m. EST, and both APIs remain fully supported until then.","note":"Also: 'Any data in your account is permanently deleted after November 13, 2026.'"},"en_gb_voices":{"value":null,"source":"https://dev.hume.ai/docs/text-to-speech-tts/overview","read":"2026-10-29","quote":"","note":"Docs page does not mention UK English voices."},"languages":{"value":"Octave 2: English, Japanese, Korean, Spanish, French, Portuguese, Italian, German, Russian, Hindi, Arabic, Octave 1: English, Spanish","source":"https://dev.hume.ai/docs/text-to-speech-tts/overview","read":"2026-10-29","quote":"Supported languages English, Spanish English, Japanese, Korean, Spanish, French, Portuguese, Italian, German, Russian, Hindi, Arabic","note":"The page prints two lists, the first (English, Spanish) is Octave 1 and the second is Octave 2."},"data_region":{"value":null,"source":"https://dev.hume.ai/docs/text-to-speech-tts/overview","read":"2026-10-29","quote":"","note":"Not mentioned on the TTS docs page."},"compliance":{"value":null,"source":"https://dev.hume.ai/docs/text-to-speech-tts/overview","read":"2026-10-29","quote":"","note":"Not mentioned on the TTS docs page."}},"terms":{"verdict":"restricts publishing","source":"https://www.hume.ai/terms-of-use","read":"2026-10-29","quote":"Except and solely to the extent such a restriction is impermissible under applicable law, Customer will not (and will not permit its users or anyone else to) do any of the following: ... (g) publish benchmarks or performance information about the Service;"}},{"id":"inworld-tts-2-flash","company":"Inworld","product":"Realtime TTS-2 Flash","layer":"tts","homepage":"https://inworld.ai/","summary":"Inworld's lower-priced speech model for live agents, aimed at low delay.","open_weights":false,"facts":{"price":{"value":"$15 per 1M characters on demand","amount":15,"currency":"USD","per":"1m_characters","tier":"on-demand","source":"https://inworld.ai/pricing","read":"2026-10-29","quote":"TTS-2 Flash $15/1M chars","note":"Pay as you go, on-demand tier. Falls to $10 (Creator, $25 a month) down to $7 (Growth, $1,500 a month), Enterprise from under $5."},"latency":{"value":null,"source":"https://inworld.ai/tts","read":"2026-10-29","quote":"","note":"The product page quotes third-party boards (Coval: 'as low as 50 ms', Cekura: 45 ms median streaming latency) rather than a vendor-measured figure, so left for the independent researcher. Docs say 'stream the first audio in tens of milliseconds' with no method."},"en_gb_voices":{"value":null,"source":"https://inworld.ai/tts","read":"2026-10-29","quote":"","note":"No UK English voice is labelled in the catalogue text read, locale counts not published on the pages checked."},"languages":{"value":"about 500 languages and dialects (product page), 200+ (docs)","source":"https://inworld.ai/tts","read":"2026-10-29","quote":"Realtime TTS-2 speaks about 500 languages","note":"The product page gives about 500 languages for the TTS-2 family and the docs say 200+, the two Inworld pages disagree on the number."},"data_region":{"value":"EU and India data residency on Enterprise only","source":"https://inworld.ai/pricing","read":"2026-10-29","quote":"EU & India data residency","note":"Page reader summary, default region for other plans not stated."},"compliance":{"value":"GDPR and SOC 2 Type II on every plan, Zero Data Retention available","source":"https://inworld.ai/pricing","read":"2026-10-29","quote":"GDPR & SOC 2 Type II on every plan","note":"HIPAA and BAA and zero data retention are add-ons on the Growth plan. SLA and DPA are Enterprise only."}},"terms":{"verdict":"restricts other use","source":"https://inworld.ai/terms","read":"2026-10-29","quote":"6.5 Benchmarking. Customer will not publicly disclose benchmarking, performance, capacity, or comparative testing results relating to the Managed GPU Services without Inworld AI's prior written consent, except as required by law. Internal benchmarking conducted by Customer for its own evaluation and procurement purposes is permitted, provided that the results are kept confidential and are not disclosed to third parties other than under written confidentiality obligations no less protective than the MSA."}},{"id":"inworld-tts-2","company":"Inworld","product":"Realtime TTS-2","layer":"tts","homepage":"https://inworld.ai/","summary":"Inworld's flagship expressive speech model.","open_weights":false,"facts":{"price":{"value":"$25 per 1M characters on demand","amount":25,"currency":"USD","per":"1m_characters","tier":"on-demand","source":"https://inworld.ai/pricing","read":"2026-10-29","quote":"TTS-2 $25/1M chars","note":"Pay as you go, on-demand tier. $20 (Creator), $17.50 (Builder), $15 (Developer), $12.50 (Growth), as low as $5 on Enterprise."},"latency":{"value":null,"source":"https://inworld.ai/tts","read":"2026-10-29","quote":"","note":"Product page quotes third-party figures only (Coval 'as low as 90 ms' at P90, Cekura 127 ms median). Left for the independent researcher."},"en_gb_voices":{"value":null,"source":"https://inworld.ai/tts","read":"2026-10-29","quote":"","note":"No UK English voice is labelled in the catalogue text read, locale counts not published on the pages checked."},"languages":{"value":"about 500 languages and dialects","source":"https://inworld.ai/tts","read":"2026-10-29","quote":"Realtime TTS-2 speaks about 500 languages","note":""},"data_region":{"value":"EU and India data residency on Enterprise only","source":"https://inworld.ai/pricing","read":"2026-10-29","quote":"EU & India data residency","note":"Page reader summary, default region for other plans not stated."},"compliance":{"value":"GDPR and SOC 2 Type II on every plan, Zero Data Retention available","source":"https://inworld.ai/pricing","read":"2026-10-29","quote":"GDPR & SOC 2 Type II on every plan","note":"HIPAA and BAA and zero data retention are add-ons on the Growth plan. SLA and DPA are Enterprise only."}},"terms":{"verdict":"restricts other use","source":"https://inworld.ai/terms","read":"2026-10-29","quote":"6.5 Benchmarking. Customer will not publicly disclose benchmarking, performance, capacity, or comparative testing results relating to the Managed GPU Services without Inworld AI's prior written consent, except as required by law. Internal benchmarking conducted by Customer for its own evaluation and procurement purposes is permitted, provided that the results are kept confidential and are not disclosed to third parties other than under written confidentiality obligations no less protective than the MSA."}},{"id":"azure-speech-neural","company":"Microsoft Azure","product":"Azure Speech neural and HD voices","layer":"tts","homepage":"https://azure.microsoft.com/en-gb/products/ai-services/text-to-speech","summary":"Microsoft's cloud text-to-speech with standard neural voices and a higher-quality HD tier.","open_weights":false,"facts":{"price":{"value":"$15 per 1M characters (neural), $22 per 1M characters (neural HD)","amount":15,"currency":"USD","per":"1m_characters","tier":"pay as you go, neural (standard); HD is $22 per 1M","source":"https://learn.microsoft.com/en-us/azure/ai-services/speech-service/speech-services-quotas-and-limits","read":"2026-10-29","quote":"Multiply the result by the unit price of $15 per million characters to estimate the monthly cost.","note":"The azure.microsoft.com pricing page loads its prices with scripts, so every figure printed as '$-' to a fetch. Prices taken from Microsoft's public Retail Prices API (UK South, USD, effective 2026-09-01 for HD): 'S1 Neural Text To Speech Characters' 15.0 per 1M and 'Neural HD Text to Speech Characters' 22.0 per 1M. Free tier 0.5M characters a month (pricing page structure). Commitment tiers exist (80M, 400M, 2,000M, 4,000M characters a month)."},"latency":{"value":"Less than 300 ms","source":"https://learn.microsoft.com/en-us/azure/ai-services/speech-service/high-definition-voices","read":"2026-10-29","quote":"Latency Less than 300 ms Greater than 500 ms Less than 300 ms","note":"Docs comparison table: Azure Speech HD voices under 300 ms and non-HD Azure Speech voices under 300 ms, Azure OpenAI HD voices over 500 ms. No method given."},"en_gb_voices":{"value":"20 (14 standard neural, 2 multilingual, 4 HD)","source":"https://learn.microsoft.com/en-us/azure/ai-services/speech-service/language-support?tabs=tts","read":"2026-10-29","quote":"en-GB-SoniaNeural","note":"Count from the en-GB rows of the language support table as read from the page (e.g. en-GB-SoniaNeural, en-GB-RyanNeural, en-GB-Ada:DragonHDLatestNeural). The page is long and only the first 100,000 characters were read."},"data_region":{"value":"UK South and UK West for neural voices, HD voices not in UK regions","source":"https://learn.microsoft.com/en-us/azure/ai-services/speech-service/regions","read":"2026-10-29","quote":"uksouth | ✅ |  | ✅ |  |  | ✅ | ✅ | ✅ | ✅","note":"Text to speech region table: uksouth has neural voices (yes), HD voices (blank). HD voices are listed for westeurope, swedencentral and francecentral among EU regions. Docs: Azure Speech stores and processes speech data in the region where the resource is created."},"compliance":{"value":null,"source":"https://learn.microsoft.com/en-us/azure/foundry/responsible-ai/speech-service/text-to-speech/data-privacy-security","read":"2026-10-29","quote":"Neither input text nor output audio content is stored in Microsoft logs.","note":"No certification list (SOC, ISO, HIPAA, GDPR) on the Speech TTS pages opened. Real-time synthesis text and audio are not stored, batch scripts are stored in Azure storage."}},"terms":{"verdict":"restricts other use","source":"https://www.microsoft.com/licensing/terms/productoffering/MicrosoftAzure/MCA","read":"2026-10-29","quote":"If Customer offers a product or service competitive to an Online Service, by using the Online Service, Customer waives any restrictions on competitive use and benchmark testing in the terms governing its competitive products and services. If Customer offers a product or service competitive to an Online Service and discloses, directly or through third parties, any benchmarks or comparative tests or evaluations (each, a “Benchmark”) of any Online Service, Customer will, upon request from Microsoft, provide: (a) all information necessary to replicate such Benchmark; and (b) access to Customer's competitive products and services for Microsoft, directly or through third parties, to perform and disclose Benchmarks. If Customer does not intend to waive such restrictions in its terms of use and agree to these terms, Customer is not allowed to use the Online Service."}},{"id":"google-chirp-3-hd","company":"Google Cloud","product":"Chirp 3: HD voices","layer":"tts","homepage":"https://cloud.google.com/text-to-speech","summary":"Google Cloud's latest high-quality text-to-speech voices, with streaming.","open_weights":false,"facts":{"price":{"value":"US$30 per 1 million characters","amount":30,"currency":"USD","per":"1m_characters","tier":"pay as you go, after free tier","source":"https://cloud.google.com/text-to-speech/pricing","read":"2026-10-29","quote":"US$0.00003 per character (US$30 per 1 million characters)","note":"Free usage limit 0 to 1 million characters a month. Instant custom voice is US$60 per 1M. Older voice types are priced lower on the same page."},"latency":{"value":null,"source":"https://docs.cloud.google.com/text-to-speech/docs/chirp3-hd","read":"2026-10-29","quote":"","note":"No latency figure on the pricing or Chirp 3 HD pages. Streaming synthesis is supported."},"en_gb_voices":{"value":"30","source":"https://docs.cloud.google.com/text-to-speech/docs/list-voices-and-types","read":"2026-10-29","quote":"English (UK) Premium en-GB en-GB-Chirp3-HD-Achernar FEMALE","note":"Count of en-GB Chirp 3 HD voices from the voices list as read from the page (30 named, e.g. Achernar ... Zubenelgenubi). The Chirp 3 HD page itself lists 30 voice names and says en-GB is a supported language."},"data_region":{"value":"EU endpoint and London (europe-west2) listed as generally available","source":"https://docs.cloud.google.com/text-to-speech/docs/chirp3-hd","read":"2026-10-29","quote":"global GA us GA eu GA asia-southeast1 GA europe-west2 GA asia-northeast1 GA","note":"Endpoints table on the Chirp 3 HD page, eu and europe-west2 (London) are listed as GA."},"compliance":{"value":"No TTS-specific certification list found, Google says it does not log customer TTS text or audio","source":"https://docs.cloud.google.com/text-to-speech/docs/data-logging","read":"2026-10-29","quote":"Google does not log any customer Cloud TTS text or audio data.","note":"Google's compliance offerings page does not name Cloud Text-to-Speech. Not confirmed here whether TTS is in scope for ISO 27001, SOC 2 or HIPAA."}},"terms":{"verdict":"restricts publishing","source":"https://cloud.google.com/terms/service-terms","read":"2026-10-29","quote":"Benchmarking. Customer may itself (but may not permit a third party to): (a) conduct benchmark tests of the Services (each a “Test”); and (b) publicly disclose the results of such Tests only if (i) the public disclosure includes all necessary information to replicate the Tests, and (ii) Customer allows Google to conduct benchmark tests of Customer's publicly available products or services and publicly disclose the results of such tests. Notwithstanding the foregoing, Customer may not do either of the following on behalf of a hyperscale public cloud provider without Google's prior written consent: (A) conduct (directly or through a third party) any Test or (B) disclose the results of any such Test."}},{"id":"amazon-polly-generative","company":"Amazon Web Services","product":"Amazon Polly (generative and neural voices)","layer":"tts","homepage":"https://aws.amazon.com/polly/","summary":"Amazon's text-to-speech service, the generative engine is its most natural voice type.","open_weights":false,"facts":{"price":{"value":"$30 per 1M characters (generative)","amount":30,"currency":"USD","per":"1m_characters","tier":"generative engine, pay as you go","source":"https://aws.amazon.com/polly/pricing/","read":"2026-10-29","quote":"priced at $30 per 1 million characters for speech requests","note":"Other engines on the same page: Neural $16.00, Standard $4.00, Long-Form $100.00 per 1M characters. Free tier: generative 100 thousand characters a month for the first 12 months, neural 1 million for 12 months, standard 5 million."},"latency":{"value":null,"source":"https://aws.amazon.com/polly/faqs/","read":"2026-10-29","quote":"you can stream audio to your users in near real time","note":"No millisecond figure on the Polly FAQ, features or pricing pages. Generative engine has a bidirectional streaming API."},"en_gb_voices":{"value":"4 (Amy, Emma, Brian, Arthur), 2 of them generative (Amy, Brian)","source":"https://docs.aws.amazon.com/polly/latest/dg/available-voices.html","read":"2026-10-29","quote":"English (British) en-GB Amy** Emma Brian Arthur Female Female Male Male","note":"All four have a neural voice. Generative: Amy and Brian. Standard: Amy, Emma, Brian. There is also a Welsh-accented English voice, Geraint (standard only). Quote condenses the table row."},"data_region":{"value":"Generative voices available in London (eu-west-2), Frankfurt and Zurich","source":"https://docs.aws.amazon.com/polly/latest/dg/generative-voices.html","read":"2026-10-29","quote":"Europe (London): eu-west-2","note":"London supports only a subset of generative voices, which includes Amy (en-GB) and Brian (en-GB). FAQ: some processed content may be stored in another AWS region for service improvement unless you opt out through AWS Support."},"compliance":{"value":"HIPAA eligible under the AWS BAA","source":"https://aws.amazon.com/polly/faqs/","read":"2026-10-29","quote":"Amazon Polly is a HIPAA Eligible Service covered under the AWS Business Associate Addendum (AWS BAA).","note":"GDPR not mentioned on the FAQ page."}},"terms":{"verdict":"restricts publishing","source":"https://aws.amazon.com/service-terms/","read":"2026-10-29","quote":"1.8. You may perform benchmarks or comparative tests or evaluations (each, a “Benchmark”) of the Services. If you perform or disclose, or direct or permit any third party to perform or disclose, any Benchmark of any of the Services, you (i) will include in any disclosure, and will disclose to us, all information necessary to replicate such Benchmark, and (ii) agree that we may perform and disclose the results of Benchmarks of your products or services, irrespective of any restrictions on Benchmarks in the terms governing your products or services."}},{"id":"resemble-ai-tts","company":"Resemble AI","product":"Resemble text-to-speech API","layer":"tts","homepage":"https://www.resemble.ai/","summary":"Resemble's hosted text-to-speech API. Resemble has stopped selling it as a commercial product, so it cannot be bought.","open_weights":false,"facts":{"price":{"value":null,"source":"https://www.resemble.ai/products/text-to-speech","read":"2026-10-29","quote":"We’ve stopped offering voice generation as a commercial product","note":"Page title block says: 'This product is no longer available'. Resemble now focuses on AI detection, watermarking and multimodal security. resemble.ai/pricing lists detection plans only."},"latency":{"value":null,"source":"https://www.resemble.ai/products/text-to-speech","read":"2026-10-29","quote":"","note":"Product withdrawn, no figures on the page. The Chatterbox GitHub page claims 'sub 200ms' for Resemble's hosted service, which is no longer sold."},"data_region":{"value":null,"source":"https://www.resemble.ai/products/text-to-speech","read":"2026-10-29","quote":"","note":"Product withdrawn."},"compliance":{"value":null,"source":"https://www.resemble.ai/pricing/","read":"2026-10-29","quote":"SOC 2 Documentation","note":"Appears in a plan comparison table for the detection product, not TTS."}},"terms":{"verdict":"restricts publishing","source":"https://www.resemble.ai/terms-of-service","read":"2026-10-29","quote":"You agree that you will not publish or make available any Content that, or use the Services, any AI Model or any detection result in a manner that: ... accesses the Services for the purposes of monitoring its availability, performance or functionality, or for any other benchmarking or competitive purposes;"}},{"id":"kokoro-82m","company":"Hexgrad (community)","product":"Kokoro-82M","layer":"tts","homepage":"https://huggingface.co/hexgrad/Kokoro-82M","summary":"A small open-weight speech model you run yourself.","open_weights":true,"facts":{"price":{"value":null,"currency":null,"per":null,"source":"https://huggingface.co/hexgrad/Kokoro-82M","read":"2026-10-29","quote":"","note":"free to download, you pay for hosting"},"licence":{"value":"Apache-2.0","source":"https://huggingface.co/hexgrad/Kokoro-82M","read":"2026-10-29","quote":"License: apache-2.0","note":""},"parameters":{"value":"82 million","source":"https://huggingface.co/hexgrad/Kokoro-82M","read":"2026-10-29","quote":"Kokoro is an open-weight TTS model with 82 million parameters.","note":""},"latency":{"value":null,"source":"https://huggingface.co/hexgrad/Kokoro-82M","read":"2026-10-29","quote":"","note":"No latency figure on the model card, depends on your hardware."},"en_gb_voices":{"value":"8 (4 female, 4 male)","source":"https://huggingface.co/hexgrad/Kokoro-82M/blob/main/VOICES.md","read":"2026-10-29","quote":"bm_george 🚹 B MM minutes C f1bc8122","note":"bf_alice, bf_emma, bf_isabella, bf_lily, bm_daniel, bm_fable, bm_george, bm_lewis. Quote condenses the table row."},"data_region":{"value":null,"source":"https://huggingface.co/hexgrad/Kokoro-82M","read":"2026-10-29","quote":"","note":"The page does not state this in words. Our reading, not a quote: Wherever you host it. Self-hosted, so data stays where you run it. No vendor statement needed."},"compliance":{"value":null,"source":"https://huggingface.co/hexgrad/Kokoro-82M","read":"2026-10-29","quote":"","note":"No compliance claims, you carry them."}},"independent":[{"by":"T5Gemma-TTS technical report (arXiv 2604.01760)","metric":"predicted naturalness UTMOS22 on LibriSpeech English, out of 5, higher is better","value":"4.51","source":"https://arxiv.org/abs/2604.01760","quote":"UTMOS↑ | 2.85±0.13 | 3.80±0.08 | 4.32±0.04 | 4.51±0.01 | 4.01±0.11","note":"Table 2, LibriSpeech (EN). Columns F5-TTS, XTTS v2, CosyVoice 2, Kokoro, T5Gemma-TTS. Kokoro uses fixed preset voices.","read":"2026-10-30"}]},{"id":"orpheus-tts","company":"Canopy Labs","product":"Orpheus TTS","layer":"tts","homepage":"https://github.com/canopyai/Orpheus-TTS","summary":"An open-weight speech model built on a Llama language model, run by you.","open_weights":true,"facts":{"price":{"value":null,"currency":null,"per":null,"source":"https://github.com/canopyai/Orpheus-TTS","read":"2026-10-29","quote":"","note":"free to download, you pay for hosting"},"licence":{"value":"Apache-2.0","source":"https://github.com/canopyai/Orpheus-TTS","read":"2026-10-29","quote":"Apache-2.0 license","note":"Repository licence. Smaller 1b, 400m and 150m sizes were listed as planned, not released."},"parameters":{"value":"Llama-3b backbone","source":"https://github.com/canopyai/Orpheus-TTS","read":"2026-10-29","quote":"built on the Llama-3b backbone","note":""},"latency":{"value":"~200ms streaming, reducible to ~100ms with input streaming","source":"https://github.com/canopyai/Orpheus-TTS","read":"2026-10-29","quote":"~200ms streaming latency for realtime applications, reducible to ~100ms with input streaming","note":"Vendor claim in the README, hardware not stated in the extract."},"en_gb_voices":{"value":"no British voices listed","source":"https://github.com/canopyai/Orpheus-TTS","read":"2026-10-29","quote":"\"tara\", \"leah\", \"jess\", \"leo\", \"dan\", \"mia\", \"zac\", \"zoe\" for English","note":"Eight English voices named, none labelled British."},"data_region":{"value":null,"source":"https://github.com/canopyai/Orpheus-TTS","read":"2026-10-29","quote":"","note":"The page does not state this in words. Our reading, not a quote: Wherever you host it. Self-hosted."},"compliance":{"value":null,"source":"https://github.com/canopyai/Orpheus-TTS","read":"2026-10-29","quote":"","note":"No compliance claims."}}},{"id":"chatterbox","company":"Resemble AI (open source)","product":"Chatterbox (Turbo, Multilingual, Nano)","layer":"tts","homepage":"https://github.com/resemble-ai/chatterbox","summary":"Resemble's open-weight speech models, free to download.","open_weights":true,"facts":{"price":{"value":null,"currency":null,"per":null,"source":"https://github.com/resemble-ai/chatterbox","read":"2026-10-29","quote":"","note":"free to download, you pay for hosting"},"licence":{"value":"MIT","source":"https://github.com/resemble-ai/chatterbox","read":"2026-10-29","quote":"MIT license","note":"Repository licence as shown on the GitHub page. Licence file itself not opened."},"parameters":{"value":"Turbo 350M, Nano 110M, Multilingual V3 and original 500M","source":"https://github.com/resemble-ai/chatterbox","read":"2026-10-29","quote":"streamlined 350M parameter architecture","note":""},"latency":{"value":"sub 200ms for Resemble's hosted service (not the open models)","source":"https://github.com/resemble-ai/chatterbox","read":"2026-10-29","quote":"ultra-low latency of sub 200ms","note":"README says this refers to the hosted service, which Resemble has since withdrawn. For Nano: 'runs 3x faster than realtime on 8 CPU cores'."},"en_gb_voices":{"value":null,"source":"https://github.com/resemble-ai/chatterbox","read":"2026-10-29","quote":"","note":"No stock British voice list on the GitHub page, not checked whether voice cloning from a reference clip is the intended route."},"languages":{"value":"English only (Turbo, Nano, original), 23+ in Multilingual V3","source":"https://github.com/resemble-ai/chatterbox","read":"2026-10-29","quote":"Chatterbox-Multilingual V3 (Language list) 500M 23+","note":""},"data_region":{"value":null,"source":"https://github.com/resemble-ai/chatterbox","read":"2026-10-29","quote":"","note":"The page does not state this in words. Our reading, not a quote: Wherever you host it. Self-hosted."},"compliance":{"value":"Every generated file carries Resemble's Perth watermark","source":"https://github.com/resemble-ai/chatterbox","read":"2026-10-29","quote":"imperceptible neural watermarks that survive MP3 compression, audio editing, and common manipulations","note":""}}},{"id":"sesame-csm-1b","company":"Sesame","product":"CSM-1B","layer":"tts","homepage":"https://github.com/SesameAILabs/csm","summary":"Sesame's open conversational speech model, run by you.","open_weights":true,"facts":{"price":{"value":null,"currency":null,"per":null,"source":"https://github.com/SesameAILabs/csm","read":"2026-10-29","quote":"","note":"free to download, you pay for hosting"},"licence":{"value":"Apache-2.0 (code), weights downloaded via Hugging Face","source":"https://github.com/SesameAILabs/csm","read":"2026-10-29","quote":"Apache-2.0 license","note":"Page says the CSM-1B checkpoint needs access through Hugging Face, separate weight terms not stated on the page. README also forbids impersonation, deceptive content such as 'fraudulent calls', and illegal use."},"parameters":{"value":"1B","source":"https://github.com/SesameAILabs/csm","read":"2026-10-29","quote":"We are releasing the 1B CSM variant.","note":""},"latency":{"value":null,"source":"https://github.com/SesameAILabs/csm","read":"2026-10-29","quote":"","note":"No latency figures, needs a CUDA GPU."},"en_gb_voices":{"value":null,"source":"https://github.com/SesameAILabs/csm","read":"2026-10-29","quote":"","note":"The FAQ says the model has some capacity for non-English languages but likely will not do well. No British voice list."},"data_region":{"value":null,"source":"https://github.com/SesameAILabs/csm","read":"2026-10-29","quote":"","note":"The page does not state this in words. Our reading, not a quote: Wherever you host it. Self-hosted."},"compliance":{"value":null,"source":"https://github.com/SesameAILabs/csm","read":"2026-10-29","quote":"","note":"No compliance claims."}}},{"id":"openai-gpt-realtime-2-1","company":"OpenAI","product":"GPT-Realtime-2.1","layer":"s2s","homepage":"https://developers.openai.com/api/docs/models/gpt-realtime-2.1","summary":"One model that listens to audio and replies in audio over the Realtime API, with tool calls and configurable reasoning.","open_weights":false,"facts":{"price_audio_in":{"value":"$32.00 per 1M audio input tokens","amount":32,"currency":"USD","per":"1m_audio_tokens_input","tier":"pay as you go, standard","source":"https://developers.openai.com/api/docs/models/gpt-realtime-2.1","read":"2026-10-29","quote":"Audio tokens Per 1M tokens Input $32.00 Cached input $0.40 Output $64.00","note":"Cached audio input is $0.4 per 1M tokens on the same page."},"price_audio_out":{"value":"$64.00 per 1M audio output tokens","amount":64,"currency":"USD","per":"1m_audio_tokens_output","tier":"pay as you go, standard","source":"https://developers.openai.com/api/docs/models/gpt-realtime-2.1","read":"2026-10-29","quote":"Audio tokens Per 1M tokens Input $32.00 Cached input $0.40 Output $64.00","note":""},"price_text_in":{"value":"$4.00 per 1M text input tokens","amount":4,"currency":"USD","per":"1m_tokens_input","tier":"pay as you go, standard","source":"https://developers.openai.com/api/docs/models/gpt-realtime-2.1","read":"2026-10-29","quote":"Text tokens Per 1M tokens Input $4.00 Cached input $0.40 Output $24.00","note":""},"price_text_out":{"value":"$24.00 per 1M text output tokens","amount":24,"currency":"USD","per":"1m_tokens_output","tier":"pay as you go, standard","source":"https://developers.openai.com/api/docs/models/gpt-realtime-2.1","read":"2026-10-29","quote":"Text tokens Per 1M tokens Input $4.00 Cached input $0.40 Output $24.00","note":""},"context_window":{"value":"128,000 tokens","source":"https://developers.openai.com/api/docs/models/gpt-realtime-2.1","read":"2026-10-29","quote":"128,000 context window","note":""},"latency":{"value":null,"source":"https://developers.openai.com/api/docs/guides/realtime","read":"2026-10-29","quote":"best starting point for voice agents that need barge-in, low first-audio latency, natural turn taking, and realtime tool use","note":"Qualitative only. No millisecond figure on the Realtime guide or the model pages."},"en_gb":{"value":null,"source":"https://developers.openai.com/api/docs/guides/voice-prompting","read":"2026-10-29","quote":"Language and accent should be controlled separately.","note":"No page lists British English as a supported accent or locale for the Realtime models. The voice prompting guide says language and accent are controlled through the prompt."},"function_calling":{"value":"yes","source":"https://developers.openai.com/api/docs/guides/realtime","read":"2026-10-29","quote":"The model works directly with audio, maintains conversation state, and can call tools.","note":"Model page lists function_calling as a supported feature."},"data_region":{"value":"US or Europe (EEA + Switzerland): storage and processing","source":"https://developers.openai.com/api/docs/guides/your-data","read":"2026-10-29","quote":"Realtime United States, Europe (EEA + Switzerland) United States, Europe (EEA + Switzerland) gpt-realtime , gpt-realtime-1.5 , gpt-realtime-mini , gpt-realtime-2 , gpt-realtime-2.1 , gpt-realtime-2.1-mini","note":"The UK has a regional storage row on the same page but its regional processing column is No. EU use requires Zero Data Retention, Modified Abuse Monitoring or similar approval on the OpenAI account (the page footnote)."},"compliance":{"value":null,"source":"https://developers.openai.com/api/docs/guides/your-data","read":"2026-10-29","quote":"","note":"No compliance certification statement (SOC 2, ISO) on the pricing, model, Realtime guide or data-residency pages that were opened."},"audio_tokens_per_minute_in":{"source":"https://developers.openai.com/api/docs/guides/voice-latency-cost","read":"2026-10-29","quote":"Audio tokens in user messages are 1 token per 100 ms of audio, while audio tokens in assistant messages are 1 token per 50ms of audio.","per":"other","value":"600 tokens a minute of caller audio (1 token per 100 ms)","amount":600,"note":"Worked out from the quoted rate: 60,000 ms at 1 token per 100 ms."},"audio_tokens_per_minute_out":{"source":"https://developers.openai.com/api/docs/guides/voice-latency-cost","read":"2026-10-29","quote":"Audio tokens in user messages are 1 token per 100 ms of audio, while audio tokens in assistant messages are 1 token per 50ms of audio.","per":"other","value":"1,200 tokens a minute of agent audio (1 token per 50 ms)","amount":1200,"note":"Worked out from the quoted rate: 60,000 ms at 1 token per 50 ms."}},"terms":{"verdict":"restricts other use","source":"https://web.archive.org/web/20261028230302/https://openai.com/policies/services-agreement/","read":"2026-10-29","quote":"You may use OpenAI models for research, software development, model evaluation and benchmarking, and product development, within these boundaries."},"independent":[{"by":"tau-Voice (arXiv 2603.13686)","model":"gpt-realtime-1.5","metric":"task completion pass@1 on realistic phone audio (noise, accents, interruptions), higher is better","value":"35%","source":"https://arxiv.org/abs/2603.13686","quote":"OpenAI 49% (-36, -42.1%) 35% (-49, -58.4%)","note":"Tested gpt-realtime-1.5, an earlier version than this row's product. Table 6, all three domains (retail, airline, telecom), 8 kHz telephony with simulated callers. Higher is better.","read":"2026-10-30"},{"by":"tau-Voice (arXiv 2603.13686)","model":"gpt-realtime-1.5","metric":"task completion pass@1 on clean phone audio, higher is better","value":"49%","source":"https://arxiv.org/abs/2603.13686","quote":"OpenAI 49% (-36, -42.1%) 35% (-49, -58.4%)","note":"Tested gpt-realtime-1.5, an earlier version than this row's product. Table 6, all three domains (retail, airline, telecom), 8 kHz telephony with simulated callers. Higher is better.","read":"2026-10-30"}]},{"id":"openai-gpt-realtime-2-1-mini","company":"OpenAI","product":"GPT-Realtime-2.1 Mini","layer":"s2s","homepage":"https://developers.openai.com/api/docs/models/gpt-realtime-2.1-mini","summary":"The cheaper, faster version of the OpenAI Realtime voice model, with audio in and audio out.","open_weights":false,"facts":{"price_audio_in":{"value":"$10.00 per 1M audio input tokens","amount":10,"currency":"USD","per":"1m_audio_tokens_input","tier":"pay as you go, standard","source":"https://developers.openai.com/api/docs/models/gpt-realtime-2.1-mini","read":"2026-10-29","quote":"Audio tokens Per 1M tokens Input $10.00 Cached input $0.30 Output $20.00","note":"Cached audio input is $0.3 per 1M tokens on the same page."},"price_audio_out":{"value":"$20.00 per 1M audio output tokens","amount":20,"currency":"USD","per":"1m_audio_tokens_output","tier":"pay as you go, standard","source":"https://developers.openai.com/api/docs/models/gpt-realtime-2.1-mini","read":"2026-10-29","quote":"Audio tokens Per 1M tokens Input $10.00 Cached input $0.30 Output $20.00","note":""},"price_text_in":{"value":"$0.60 per 1M text input tokens","amount":0.6,"currency":"USD","per":"1m_tokens_input","tier":"pay as you go, standard","source":"https://developers.openai.com/api/docs/models/gpt-realtime-2.1-mini","read":"2026-10-29","quote":"Text tokens Per 1M tokens Input $0.60 Cached input $0.06 Output $2.40","note":""},"price_text_out":{"value":"$2.40 per 1M text output tokens","amount":2.4,"currency":"USD","per":"1m_tokens_output","tier":"pay as you go, standard","source":"https://developers.openai.com/api/docs/models/gpt-realtime-2.1-mini","read":"2026-10-29","quote":"Text tokens Per 1M tokens Input $0.60 Cached input $0.06 Output $2.40","note":""},"context_window":{"value":"128,000 tokens","source":"https://developers.openai.com/api/docs/models/gpt-realtime-2.1-mini","read":"2026-10-29","quote":"128,000 context window","note":""},"latency":{"value":null,"source":"https://developers.openai.com/api/docs/guides/realtime","read":"2026-10-29","quote":"best starting point for voice agents that need barge-in, low first-audio latency, natural turn taking, and realtime tool use","note":"Qualitative only. No millisecond figure on the Realtime guide or the model pages."},"en_gb":{"value":null,"source":"https://developers.openai.com/api/docs/guides/voice-prompting","read":"2026-10-29","quote":"Language and accent should be controlled separately.","note":"No page lists British English as a supported accent or locale for the Realtime models. The voice prompting guide says language and accent are controlled through the prompt."},"function_calling":{"value":"yes","source":"https://developers.openai.com/api/docs/guides/realtime","read":"2026-10-29","quote":"The model works directly with audio, maintains conversation state, and can call tools.","note":"Model page lists function_calling as a supported feature."},"data_region":{"value":"US or Europe (EEA + Switzerland): storage and processing","source":"https://developers.openai.com/api/docs/guides/your-data","read":"2026-10-29","quote":"Realtime United States, Europe (EEA + Switzerland) United States, Europe (EEA + Switzerland) gpt-realtime , gpt-realtime-1.5 , gpt-realtime-mini , gpt-realtime-2 , gpt-realtime-2.1 , gpt-realtime-2.1-mini","note":"The UK has a regional storage row on the same page but its regional processing column is No. EU use requires Zero Data Retention, Modified Abuse Monitoring or similar approval on the OpenAI account (the page footnote)."},"compliance":{"value":null,"source":"https://developers.openai.com/api/docs/guides/your-data","read":"2026-10-29","quote":"","note":"No compliance certification statement (SOC 2, ISO) on the pricing, model, Realtime guide or data-residency pages that were opened."},"audio_tokens_per_minute_in":{"source":"https://developers.openai.com/api/docs/guides/voice-latency-cost","read":"2026-10-29","quote":"Audio tokens in user messages are 1 token per 100 ms of audio, while audio tokens in assistant messages are 1 token per 50ms of audio.","per":"other","value":"600 tokens a minute of caller audio (1 token per 100 ms)","amount":600,"note":"Worked out from the quoted rate: 60,000 ms at 1 token per 100 ms."},"audio_tokens_per_minute_out":{"source":"https://developers.openai.com/api/docs/guides/voice-latency-cost","read":"2026-10-29","quote":"Audio tokens in user messages are 1 token per 100 ms of audio, while audio tokens in assistant messages are 1 token per 50ms of audio.","per":"other","value":"1,200 tokens a minute of agent audio (1 token per 50 ms)","amount":1200,"note":"Worked out from the quoted rate: 60,000 ms at 1 token per 50 ms."}},"terms":{"verdict":"restricts other use","source":"https://web.archive.org/web/20261028230302/https://openai.com/policies/services-agreement/","read":"2026-10-29","quote":"You may use OpenAI models for research, software development, model evaluation and benchmarking, and product development, within these boundaries."}},{"id":"openai-gpt-live-1","company":"OpenAI","product":"GPT-Live 1","layer":"s2s","homepage":"https://developers.openai.com/api/docs/models/gpt-live-1","summary":"A voice model that listens and talks at the same time, billed per minute of conversation, and hands reasoning and tool use to a separate backend model.","open_weights":false,"facts":{"price":{"value":"$0.05 per minute of voice session, billed per second","amount":0.05,"currency":"USD","per":"call_minute","tier":"pay as you go; backend model and tool tokens extra","source":"https://developers.openai.com/api/docs/models/gpt-live-1","read":"2026-10-29","quote":"Voice sessions cost $0.05 per minute, billed per second. Backend model and tool usage is billed separately.","note":"Not token-priced. The backend model that does the reasoning is billed at its own rate, so the total per minute is higher than $0.05."},"function_calling":{"value":"yes, through a delegated backend agent","source":"https://developers.openai.com/api/docs/models/gpt-live-1","read":"2026-10-29","quote":"It can listen and speak at the same time, and delegate reasoning and tool use to a backend agent.","note":"function_calling is listed under supported features."},"latency":{"value":null,"source":"https://developers.openai.com/api/docs/models/gpt-live-1","read":"2026-10-29","quote":"Our premier model for natural, expressive voice conversations with smooth interruption handling.","note":"No latency figure on the model page."},"en_gb":{"value":null,"source":"https://developers.openai.com/api/docs/models/gpt-live-1","read":"2026-10-29","quote":"","note":"No language or accent list on the model page."},"data_region":{"value":null,"source":"https://developers.openai.com/api/docs/guides/your-data","read":"2026-10-29","quote":"","note":"The data-residency page lists the /v1/live/sessions route under the US and Europe rows, the model-level list was not checked for gpt-live-1."},"compliance":{"value":null,"source":"https://developers.openai.com/api/docs/guides/your-data","read":"2026-10-29","quote":"","note":"No compliance certification statement (SOC 2, ISO) on the pricing, model, Realtime guide or data-residency pages that were opened."},"price_audio_in":{"value":null,"source":"https://developers.openai.com/api/docs/models/gpt-live-1","read":"2026-10-29","quote":"","note":"Not priced per audio token. See the price fact for the per-minute rate."},"price_audio_out":{"value":null,"source":"https://developers.openai.com/api/docs/models/gpt-live-1","read":"2026-10-29","quote":"","note":"Not priced per audio token. See the price fact for the per-minute rate."}},"terms":{"verdict":"restricts other use","source":"https://web.archive.org/web/20261028230302/https://openai.com/policies/services-agreement/","read":"2026-10-29","quote":"You may use OpenAI models for research, software development, model evaluation and benchmarking, and product development, within these boundaries."}},{"id":"google-gemini-3-8-live","company":"Google","product":"Gemini 3.8 Live (Live API)","layer":"s2s","homepage":"https://ai.google.dev/gemini-api/docs/live-api","summary":"Google's audio-in, audio-out model for voice agents, used through the Gemini Live API.","open_weights":false,"facts":{"price_audio_in":{"value":"$3.00 per 1M audio input tokens, or $0.005 per minute","amount":3,"currency":"USD","per":"1m_audio_tokens_input","tier":"paid tier","amount_per_minute":0.005,"source":"https://ai.google.dev/gemini-api/docs/pricing?hl=en","read":"2026-10-29","quote":"Input price Free of charge $0.75 (text) $3.00 or $0.005/min (audio) $1.00 or $0.002/min (image/video)","note":"Same price row covers gemini-3.8-live, gemini-3.8-live-extended-thinking and gemini-3.1-flash-live-preview. Per-minute figure is Google's own."},"price_audio_out":{"value":"$12.00 per 1M audio output tokens, or $0.018 per minute","amount":12,"currency":"USD","per":"1m_audio_tokens_output","tier":"paid tier","amount_per_minute":0.018,"source":"https://ai.google.dev/gemini-api/docs/pricing?hl=en","read":"2026-10-29","quote":"Output price (including thinking tokens) Free of charge $4.50 (text) $12.00 or $0.018/min (audio)","note":""},"price_text_in":{"value":"$0.75 per 1M text input tokens","amount":0.75,"currency":"USD","per":"1m_tokens_input","tier":"paid tier","source":"https://ai.google.dev/gemini-api/docs/pricing?hl=en","read":"2026-10-29","quote":"Input price Free of charge $0.75 (text)","note":""},"price_text_out":{"value":"$4.50 per 1M text output tokens","amount":4.5,"currency":"USD","per":"1m_tokens_output","tier":"paid tier","source":"https://ai.google.dev/gemini-api/docs/pricing?hl=en","read":"2026-10-29","quote":"Output price (including thinking tokens) Free of charge $4.50 (text)","note":""},"tokens_per_minute":{"value":null,"source":"https://ai.google.dev/gemini-api/docs/pricing?hl=en","read":"2026-10-29","quote":"","note":"Google prints per-minute prices for Live audio but no tokens-per-minute figure for this model. The 25 tokens per second figure on the same page is stated only for Live Translate and Live Transcribe."},"latency":{"value":null,"source":"https://ai.google.dev/gemini-api/docs/models/gemini-3.8-live?hl=en","read":"2026-10-29","quote":"Gemini 3.8 Live delivers ultra-low latency audio-to-audio interactions","note":"Qualitative only. No millisecond figure on the model page."},"en_gb":{"value":null,"source":"https://ai.google.dev/gemini-api/docs/live-guide?hl=en","read":"2026-10-29","quote":"Live API supports the following 99 languages.","note":"The language table lists English as \"en\" with no regional variants."},"function_calling":{"value":"yes, asynchronous by default, blocking mode still available","source":"https://ai.google.dev/gemini-api/docs/models/gemini-3.8-live?hl=en","read":"2026-10-29","quote":"Asynchronous function calling : Async execution ( behavior: NON_BLOCKING ) is now the default function calling mode.","note":""},"data_region":{"value":null,"source":"https://ai.google.dev/gemini-api/docs/pricing?hl=en","read":"2026-10-29","quote":"","note":"No region or residency statement on the pricing, model or Live API pages that were opened."},"compliance":{"value":null,"source":"https://ai.google.dev/gemini-api/docs/pricing?hl=en","read":"2026-10-29","quote":"","note":"No compliance statement on the pricing, model or Live API pages that were opened."}},"terms":{"verdict":"restricts other use","source":"https://ai.google.dev/gemini-api/terms","read":"2026-10-29","quote":"Customer may itself (but may not permit a third party to): (a) conduct benchmark tests of the Services (each a “Test”); and (b) publicly disclose the results of such Tests only if (i) the public disclosure includes all necessary information to replicate the Tests, and (ii) Customer allows Google to conduct benchmark tests of Customer's publicly available products or services and publicly disclose the results of such tests."},"independent":[{"by":"tau-Voice (arXiv 2603.13686)","model":"Gemini Live 2.5 Flash","metric":"task completion pass@1 on realistic phone audio (noise, accents, interruptions), higher is better","value":"26%","source":"https://arxiv.org/abs/2603.13686","quote":"Google 85% (54%) 31% (-54, -63.4%) 26% (-59, -69.5%)","note":"Tested Gemini Live 2.5 Flash, an earlier version than this row's product. Table 6, all three domains (retail, airline, telecom), 8 kHz telephony with simulated callers. Higher is better.","read":"2026-10-30"},{"by":"tau-Voice (arXiv 2603.13686)","model":"Gemini Live 2.5 Flash","metric":"task completion pass@1 on clean phone audio, higher is better","value":"31%","source":"https://arxiv.org/abs/2603.13686","quote":"Google 85% (54%) 31% (-54, -63.4%) 26% (-59, -69.5%)","note":"Tested Gemini Live 2.5 Flash, an earlier version than this row's product. Table 6, all three domains (retail, airline, telecom), 8 kHz telephony with simulated callers. Higher is better.","read":"2026-10-30"}]},{"id":"amazon-nova-2-5-sonic","company":"Amazon (AWS)","product":"Nova 2.5 Sonic","layer":"s2s","homepage":"https://aws.amazon.com/nova/","summary":"Amazon's audio-in, audio-out voice model, run through Amazon Bedrock, with text and voice allowed in the same session.","open_weights":false,"facts":{"price_audio_in":{"value":"$3.00 per 1M speech input tokens","amount":3,"currency":"USD","per":"1m_audio_tokens_input","tier":"on demand, standard tier, US East (N. Virginia)","source":"https://aws.amazon.com/nova/pricing/","read":"2026-10-29","quote":"Amazon Nova 2.5 Sonic Speech $3.00 $12.00","note":"The rendered page shows one price list for a chosen region. The price table loads by script, so the first fetch of the page showed blanks and the figures were read in a real browser. Other regions were not compared."},"price_audio_out":{"value":"$12.00 per 1M speech output tokens","amount":12,"currency":"USD","per":"1m_audio_tokens_output","tier":"on demand, standard tier, US East (N. Virginia)","source":"https://aws.amazon.com/nova/pricing/","read":"2026-10-29","quote":"Amazon Nova 2.5 Sonic Speech $3.00 $12.00","note":""},"price_text_in":{"value":"$0.33 per 1M text input tokens","amount":0.33,"currency":"USD","per":"1m_tokens_input","tier":"on demand, standard tier, US East (N. Virginia)","source":"https://aws.amazon.com/nova/pricing/","read":"2026-10-29","quote":"Amazon Nova 2.5 Sonic Text $0.33 $2.75","note":"Page note: text tokens apply to uses such as transcription, tool calls and conversation history added to the session."},"price_text_out":{"value":"$2.75 per 1M text output tokens","amount":2.75,"currency":"USD","per":"1m_tokens_output","tier":"on demand, standard tier, US East (N. Virginia)","source":"https://aws.amazon.com/nova/pricing/","read":"2026-10-29","quote":"Amazon Nova 2.5 Sonic Text $0.33 $2.75","note":""},"tokens_per_minute":{"value":null,"source":"https://aws.amazon.com/nova/pricing/","read":"2026-10-29","quote":"","note":"No tokens-per-minute figure on the pricing page, the model card or the service card that were opened."},"context_window":{"value":"256K tokens","source":"https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-amazon-nova-2-5-sonic.md","read":"2026-10-29","quote":"Context window:** 256K tokens","note":""},"latency":{"value":null,"source":"https://docs.aws.amazon.com/ai/responsible-ai/nova-2-5-sonic/overview.html","read":"2026-10-29","quote":"offers the following core capabilities: low latency, speech understanding in multiple languages","note":"Qualitative only. The service card lists low latency as a capability but prints no figure."},"en_gb":{"value":"yes","source":"https://docs.aws.amazon.com/ai/responsible-ai/nova-2-5-sonic/overview.html","read":"2026-10-29","quote":"with expressive voices optimized for US English, British English, Indian English, Australian English","note":"The Nova 2 language page lists English (UK) en-GB with one feminine voice, \"amy\", and no masculine voice. The 2.5 service card says it has the same voices and language support as Nova 2 Sonic."},"function_calling":{"value":"yes, including asynchronous tool calling","source":"https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-amazon-nova-2-5-sonic.md","read":"2026-10-29","quote":"Higher-accuracy tool calling for agentic workflows: look up an order, check eligibility, and take action in one session.","note":""},"data_region":{"value":"In-region only: N. Virginia, Oregon, Stockholm, Tokyo. No geo or global cross-region routing.","source":"https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-amazon-nova-2-5-sonic.md","read":"2026-10-29","quote":"| bedrock-runtime | amazon.nova-2-5-sonic | https://bedrock-runtime.{region}.amazonaws.com | Not supported | Not supported |","note":"Regions from the model card table: us-east-1, us-west-2, eu-north-1 (Stockholm), ap-northeast-1 (Tokyo). No London region is listed. The last two \"Not supported\" cells are the geo and global inference IDs."},"compliance":{"value":"GDPR and HIPAA named for Bedrock customers","source":"https://docs.aws.amazon.com/ai/responsible-ai/nova-2-5-sonic/overview.html","read":"2026-10-29","quote":"enterprise security that enables customers to build generative AI applications that support common data security and compliance standards, including GDPR and HIPAA","note":"The card names the standards but gives no certification list."}},"terms":{"verdict":"restricts publishing","source":"https://aws.amazon.com/service-terms/","read":"2026-10-29","quote":"1.8. You may perform benchmarks or comparative tests or evaluations (each, a “Benchmark”) of the Services. If you perform or disclose, or direct or permit any third party to perform or disclose, any Benchmark of any of the Services, you (i) will include in any disclosure, and will disclose to us, all information necessary to replicate such Benchmark, and (ii) agree that we may perform and disclose the results of Benchmarks of your products or services, irrespective of any restrictions on Benchmarks in the terms governing your products or services."}},{"id":"azure-voice-live","company":"Microsoft","product":"Azure Voice Live API","layer":"s2s","homepage":"https://learn.microsoft.com/en-us/azure/ai-services/speech-service/voice-live","summary":"Microsoft's managed voice agent service that runs OpenAI GPT-Realtime models (or text models with Azure speech) behind one WebSocket, billed per token in three tiers.","open_weights":false,"facts":{"tiers":{"value":"gpt-realtime-2.1 is billed on the Pro tier, gpt-realtime-2.1-mini on the Standard tier","source":"https://learn.microsoft.com/en-us/azure/ai-services/speech-service/voice-live","read":"2026-10-29","quote":"gpt-realtime-2.1 Pro GPT real-time 2.1 with global processing.","note":"The docs table also lists Lite (gpt-4.1-nano, gpt-6-luna). Pricing page: \"Voice Live pricing tiers are determined by the LLM models selected by developers.\""},"price_audio_in":{"value":"$32.00 per 1M audio input tokens (Pro tier, gpt-realtime-2.1)","amount":32,"currency":"USD","per":"1m_audio_tokens_input","tier":"Pro tier, US East 2 row, pay as you go","source":"https://prices.azure.com/api/retail/prices?currencyCode=%27USD%27&%24filter=serviceName%20eq%20%27Foundry%20Tools%27%20and%20contains(meterName,%27Voice%20Live%27)","read":"2026-10-29","quote":"\"retailPrice\":0.032,\"unitPrice\":0.032,\"armRegionName\":\"eastus2\",\"location\":\"US East 2\",\"effectiveStartDate\":\"2026-05-01T00:00:00Z\",\"meterId\":\"27738f15-582c-5090-b8c4-93fa0b8c9295\",\"meterName\":\"Voice Live API Pro - LLM Audio Input Tokens\"","note":"The API response had no UK South row for this meter, every region it did return (US East 2, Japan East, Sweden Central, others) shows the same 0.032, so US East 2 is quoted. Read from Microsoft's Retail Prices API as 0.032 USD per 1K tokens. The azure.microsoft.com pricing page prints \"$-\" placeholders in the fetched HTML, so the API was used."},"price_audio_out":{"value":"$64.00 per 1M audio output tokens (Pro tier, gpt-realtime-2.1)","amount":64,"currency":"USD","per":"1m_audio_tokens_output","tier":"Pro tier, UK South, pay as you go","source":"https://prices.azure.com/api/retail/prices?currencyCode=%27USD%27&%24filter=serviceName%20eq%20%27Foundry%20Tools%27%20and%20contains(meterName,%27Voice%20Live%27)","read":"2026-10-29","quote":"\"retailPrice\":0.064,\"unitPrice\":0.064,\"armRegionName\":\"uksouth\",\"location\":\"UK South\",\"effectiveStartDate\":\"2026-05-01T00:00:00Z\",\"meterId\":\"868f641f-6fee-50ac-92e3-f9acf14eb72c\",\"meterName\":\"Voice Live API Pro - LLM Audio Output Tokens\"","note":"0.064 USD per 1K tokens."},"price_text_in":{"value":"$4.00 per 1M text input tokens (Pro tier)","amount":4,"currency":"USD","per":"1m_tokens_input","tier":"Pro tier, UK South, pay as you go","source":"https://prices.azure.com/api/retail/prices?currencyCode=%27USD%27&%24filter=serviceName%20eq%20%27Foundry%20Tools%27%20and%20contains(meterName,%27Voice%20Live%27)","read":"2026-10-29","quote":"\"retailPrice\":0.004,\"unitPrice\":0.004,\"armRegionName\":\"uksouth\",\"location\":\"UK South\",\"effectiveStartDate\":\"2026-05-01T00:00:00Z\",\"meterId\":\"d461d4f9-cafc-5936-a901-0de9845453fa\",\"meterName\":\"Voice Live API Pro - LLM Text Input Tokens\"","note":"0.004 USD per 1K tokens."},"price_text_out":{"value":"$24.00 per 1M text output tokens (Pro tier, gpt-realtime-2 and 2.1)","amount":24,"currency":"USD","per":"1m_tokens_output","tier":"Pro tier, UK South, pay as you go","source":"https://prices.azure.com/api/retail/prices?currencyCode=%27USD%27&%24filter=serviceName%20eq%20%27Foundry%20Tools%27%20and%20contains(meterName,%27Voice%20Live%27)","read":"2026-10-29","quote":"\"retailPrice\":0.024,\"unitPrice\":0.024,\"armRegionName\":\"uksouth\",\"location\":\"UK South\",\"effectiveStartDate\":\"2026-09-01T00:00:00Z\",\"meterId\":\"5b56eb25-2d2f-5d31-8b1a-9153cfd8a221\",\"meterName\":\"Voice Live API Pro - LLM Text Output 2 Tokens\"","note":"0.024 USD per 1K tokens. The pricing page labels this \"Output 2 (gpt-realtime-2, gpt-realtime-2.1)\", the older \"Output 1\" meter is 0.016 per 1K."},"price_audio_in_mini":{"value":"$11.00 per 1M audio input tokens (Standard tier, gpt-realtime-2.1-mini)","amount":11,"currency":"USD","per":"1m_audio_tokens_input","tier":"Standard tier, UK South, pay as you go","source":"https://prices.azure.com/api/retail/prices?currencyCode=%27USD%27&%24filter=serviceName%20eq%20%27Foundry%20Tools%27%20and%20contains(meterName,%27Voice%20Live%27)","read":"2026-10-29","quote":"\"retailPrice\":0.011,\"unitPrice\":0.011,\"armRegionName\":\"uksouth\",\"location\":\"UK South\",\"effectiveStartDate\":\"2025-10-01T00:00:00Z\",\"meterId\":\"9da6868e-ea60-588f-8b9d-0ced508e6a9d\",\"meterName\":\"Voice Live API Std - LLM Audio Input Tokens\"","note":"0.011 USD per 1K tokens. Higher than OpenAI's own $10 for the mini model."},"price_audio_out_mini":{"value":"$22.00 per 1M audio output tokens (Standard tier, gpt-realtime-2.1-mini)","amount":22,"currency":"USD","per":"1m_audio_tokens_output","tier":"Standard tier, UK South, pay as you go","source":"https://prices.azure.com/api/retail/prices?currencyCode=%27USD%27&%24filter=serviceName%20eq%20%27Foundry%20Tools%27%20and%20contains(meterName,%27Voice%20Live%27)","read":"2026-10-29","quote":"\"retailPrice\":0.022,\"unitPrice\":0.022,\"armRegionName\":\"uksouth\",\"location\":\"UK South\",\"effectiveStartDate\":\"2025-10-01T00:00:00Z\",\"meterId\":\"c81a7df4-2ada-53ff-b03e-6d53bce38e93\",\"meterName\":\"Voice Live API Std- LLM Audio Output Tokens\"","note":"0.022 USD per 1K tokens. Meter name has a typo (\"Std-\") in the API."},"tokens_per_minute":{"value":null,"source":"https://azure.microsoft.com/en-gb/pricing/details/cognitive-services/speech-services/","read":"2026-10-29","quote":"Voice Live pricing tiers are determined by the LLM models selected by developers.","note":"No tokens-per-minute figure on the pricing page or the Voice Live docs. Microsoft passes through the OpenAI model's token accounting."},"latency":{"value":null,"source":"https://learn.microsoft.com/en-us/azure/ai-services/speech-service/voice-live","read":"2026-10-29","quote":"The Voice Live API is a solution that enables low-latency, high-quality speech-to-speech interactions for voice agents.","note":"Qualitative only. No millisecond figure on the overview page."},"en_gb":{"value":"yes","source":"https://learn.microsoft.com/en-us/azure/ai-services/speech-service/language-support?tabs=stt","read":"2026-10-29","quote":"en-GB English (United Kingdom) Neural HD en-GB-Ada:DragonHDLatestNeural 2,4 (Female)","note":"From the Azure Speech language support page: en-GB listed for speech to text and for HD voices such as en-GB-Ada. Voice Live uses Azure Speech for these when you choose an Azure voice instead of the native OpenAI audio voices."},"function_calling":{"value":"yes","source":"https://learn.microsoft.com/en-us/azure/ai-services/speech-service/voice-live","read":"2026-10-29","quote":"Function calling : Enables external actions, use of tools, and grounded responses using the VoiceRAG pattern.","note":""},"data_region":{"value":"Varies: speech, voices and the model have separate regions and processing scopes. gpt-realtime-2.1 is offered as global, data-zone or regional processing.","source":"https://learn.microsoft.com/en-us/azure/ai-services/speech-service/regions","read":"2026-10-29","quote":"Voice Live is an exception: speech recognition, HD voice synthesis, and LLM inference have separate processing scopes. Processing can occur outside your resource region.","note":"Check both the speech table and the LLM table on the Microsoft region pages before assuming data stays in the UK."},"compliance":{"value":null,"source":"https://learn.microsoft.com/en-us/azure/ai-services/speech-service/voice-live","read":"2026-10-29","quote":"","note":"No compliance statement on the Voice Live overview or pricing pages that were opened."}},"terms":{"verdict":"restricts other use","source":"https://www.microsoft.com/licensing/terms/productoffering/MicrosoftAzure/MCA","read":"2026-10-29","quote":"If Customer offers a product or service competitive to an Online Service, by using the Online Service, Customer waives any restrictions on competitive use and benchmark testing in the terms governing its competitive products and services. If Customer offers a product or service competitive to an Online Service and discloses, directly or through third parties, any benchmarks or comparative tests or evaluations (each, a “Benchmark”) of any Online Service, Customer will, upon request from Microsoft, provide: (a) all information necessary to replicate such Benchmark; and (b) access to Customer's competitive products and services for Microsoft, directly or through third parties, to perform and disclose Benchmarks. If Customer does not intend to waive such restrictions in its terms of use and agree to these terms, Customer is not allowed to use the Online Service."}},{"id":"ultravox-realtime","company":"Ultravox (Fixie.ai, acquired by Inworld)","product":"Ultravox Realtime (open-weight speech model plus hosting)","layer":"s2s","homepage":"https://www.ultravox.ai/","summary":"A hosted voice agent service built on its own open-weight model that listens to speech directly, priced per call minute, not per token.","open_weights":true,"facts":{"price":{"value":"$0.05 per minute after 30 free minutes","amount":0.05,"currency":"USD","per":"call_minute","tier":"pay as you go, up to 5 concurrent calls","source":"https://www.ultravox.ai/pricing","read":"2026-10-29","quote":"First 30 minutes free, $0.05/min after that.","note":"Calls are rounded to the nearest six seconds (FAQ). Pro plan is $100 a month with no concurrency cap and the same per-minute rate. Pay-as-you-go SIP is billed at 0.5 cents per minute on top, the page does not say what else the per-minute rate covers."},"status":{"value":"Acquired by Inworld, pricing page still live on 2026-10-29","source":"https://www.ultravox.ai/pricing","read":"2026-10-29","quote":"Ultravox has been acquired by Inworld to accelerate its speech-to-speech research and voice agent capabilities.","note":""},"tokens_per_minute":{"value":null,"source":"https://www.ultravox.ai/pricing","read":"2026-10-29","quote":"","note":"Not token-priced for voice calls. Ultravox publishes per-token prices only for its text \"Threads\" product."},"latency":{"value":null,"source":"https://docs.ultravox.ai/gettingstarted/how-ultravox-works.md","read":"2026-10-29","quote":"Voice AI only works when conversations feel natural and fluid.","note":"Docs point to a blog post for speed comparisons, no millisecond figure on the docs pages that were opened."},"en_gb":{"value":null,"source":"https://docs.ultravox.ai/gettingstarted/how-ultravox-works.md","read":"2026-10-29","quote":"","note":"No language list on the docs pages that were opened."},"function_calling":{"value":"yes","source":"https://docs.ultravox.ai/gettingstarted/how-ultravox-works.md","read":"2026-10-29","quote":"Tools (AKA function calling) give your agents superpowers","note":""},"data_region":{"value":null,"source":"https://docs.ultravox.ai/gettingstarted/faq.md","read":"2026-10-29","quote":"Call data is retained until you delete it.","note":"No region statement on the FAQ or pricing pages. Retention: call data kept until you delete it, automatic deletion is Enterprise only."},"compliance":{"value":"HIPAA BAA available on Enterprise plan","source":"https://docs.ultravox.ai/gettingstarted/faq.md","read":"2026-10-29","quote":"We sign Business Associate Agreements (BAAs) for customers with HIPAA needs, on an Enterprise plan.","note":"No SOC 2 or ISO statement on the pages opened."},"price_audio_in":{"value":null,"source":"https://www.ultravox.ai/pricing","read":"2026-10-29","quote":"","note":"Not priced per audio token. See the price fact for the per-minute rate."},"price_audio_out":{"value":null,"source":"https://www.ultravox.ai/pricing","read":"2026-10-29","quote":"","note":"Not priced per audio token. See the price fact for the per-minute rate."}},"independent":[{"by":"Full-Duplex-Bench-v3 (arXiv 2604.04847)","model":"Ultravox v0.7","metric":"pass@1 on tool-use tasks with real hesitant speech, higher is better","value":"0.410","source":"https://arxiv.org/abs/2604.04847","quote":"Ultravox | 0.794 | 0.513 | 0.510 | 0.410 | 96.0% | 8.40s | 47.9% | 88.0%","note":"Table 2, columns Tool Sel, Arg Acc, Resp Qual, Pass@1, Take-turn, Latency, Interrupt, Filler.","read":"2026-10-30"},{"by":"EVA-Bench (arXiv 2605.13841)","metric":"task accuracy pass@1 over 213 scenarios, clean audio, higher is better","value":"0.270","source":"https://arxiv.org/abs/2605.13841","quote":"Hybrid | Ultravox-Realtime | 0.270 ±0.047 | 0.503 ±0.072 | 0.108 ±0.037","note":"Table 4, pass@1 column, 5 trials. Tested as a hybrid system, version not stated.","read":"2026-10-30"}]},{"id":"hume-evi","company":"Hume AI","product":"Empathic Voice Interface (EVI)","layer":"s2s","homepage":"https://dev.hume.ai/voice/intro","summary":"Hume's voice-to-voice model for emotionally aware agents. Hume is closing this API on 13 November 2026, so it cannot be used after that date.","open_weights":false,"facts":{"status":{"value":"API shuts down 13 November 2026","source":"https://dev.hume.ai/voice/intro","read":"2026-10-29","quote":"Hume is sunsetting the TTS and EVI APIs. Access ends November 13, 2026 at 12:01 a.m. EST, and both APIs remain fully supported until then.","note":"Data in accounts is permanently deleted after that date. Not a sensible option for a new build."},"price":{"value":null,"source":"https://dev.hume.ai/docs/resources/billing","read":"2026-10-29","quote":"Overage billing applies when you exceed included usage; see plan rates at our pricing page","note":"Hume bills EVI through monthly subscriptions with overage rates, and sends you to its pricing page for the rates. That page (hume.ai/pricing) now describes only Hume's evaluation products and prints no EVI rates, so no per-minute price could be read from Hume. Third-party sites quote about $0.04 to $0.07 per minute but those are not used here."},"tokens_per_minute":{"value":null,"source":"https://dev.hume.ai/docs/resources/billing","read":"2026-10-29","quote":"","note":"Billed per minute through subscriptions, not per token."},"latency":{"value":null,"source":"https://dev.hume.ai/voice/intro","read":"2026-10-29","quote":"","note":"No latency figure on the intro page."},"en_gb":{"value":null,"source":"https://dev.hume.ai/voice/intro","read":"2026-10-29","quote":"","note":"No language or accent list on the intro page."},"function_calling":{"value":null,"source":"https://dev.hume.ai/voice/intro","read":"2026-10-29","quote":"","note":"The docs navigation lists \"Tool use\" under EVI features, the page itself was not opened, so no statement is recorded."},"data_region":{"value":null,"source":"https://dev.hume.ai/voice/intro","read":"2026-10-29","quote":"","note":"No region statement on the pages opened."},"compliance":{"value":null,"source":"https://dev.hume.ai/voice/intro","read":"2026-10-29","quote":"","note":"No compliance statement on the pages opened."},"price_audio_in":{"value":null,"source":"https://dev.hume.ai/docs/resources/billing","read":"2026-10-29","quote":"","note":"Not priced per audio token. See the price fact for the per-minute rate."},"price_audio_out":{"value":null,"source":"https://dev.hume.ai/docs/resources/billing","read":"2026-10-29","quote":"","note":"Not priced per audio token. See the price fact for the per-minute rate."}},"terms":{"verdict":"restricts publishing","source":"https://www.hume.ai/terms-of-use","read":"2026-10-29","quote":"Except and solely to the extent such a restriction is impermissible under applicable law, Customer will not (and will not permit its users or anyone else to) do any of the following: ... (g) publish benchmarks or performance information about the Service;"}},{"id":"xai-grok-voice-3-0","company":"xAI (docs now branded SpaceXSI)","product":"Grok Voice (grok-voice-3.0) speech-to-speech API","layer":"s2s","homepage":"https://docs.x.ai/developers/model-capabilities/audio/voice","summary":"xAI's realtime voice agent API over WebSockets, billed per minute of conversation.","open_weights":false,"facts":{"price":{"value":"$0.08 per minute ($4.80 per hour), plus $0.004 per text input","amount":0.08,"currency":"USD","per":"call_minute","tier":"pay as you go","source":"https://docs.x.ai/developers/models.md","read":"2026-10-29","quote":"| Speech to Speech (grok-voice-3.0) | $0.08 / min ($4.80 / hr) $0.004 / text input |","note":"The page does not say what unit \"$0.004 / text input\" is charged on. Treat the $0.004 as unclear until xAI defines it. grok-voice-2.0 appears in the same pricing data at the same rates."},"tokens_per_minute":{"value":null,"source":"https://docs.x.ai/docs/models","read":"2026-10-29","quote":"","note":"Not token-priced for voice."},"latency":{"value":null,"source":"https://docs.x.ai/developers/model-capabilities/audio/voice","read":"2026-10-29","quote":"enterprise-grade reliability and sub-second latency","note":"Qualitative only. \"Sub-second\" is the vendor's wording, no measured figure."},"en_gb":{"value":null,"source":"https://docs.x.ai/developers/model-capabilities/audio/voice","read":"2026-10-29","quote":"Each has its own personality and tone, and all of them can speak every supported language.","note":"The voice overview does not list the supported languages or accents."},"function_calling":{"value":"yes","source":"https://docs.x.ai/developers/model-capabilities/audio/voice","read":"2026-10-29","quote":"Build real-time, speech-to-speech voice agents over WebSockets, with low-latency turn-taking and tool use.","note":""},"data_region":{"value":"EU data residency options","source":"https://docs.x.ai/developers/model-capabilities/audio/voice","read":"2026-10-29","quote":"GDPR Compliant Data processing agreements and EU data residency options","note":"Also lists regional processing and multi-region infrastructure for enterprise. No UK region named."},"compliance":{"value":"SOC 2 Type II, HIPAA eligible (BAA available), GDPR, zero data retention option","source":"https://docs.x.ai/developers/model-capabilities/audio/voice","read":"2026-10-29","quote":"SOC 2 Type II Audited controls for security, availability, and confidentiality","note":"Same page lists \"HIPAA Eligible\", \"GDPR Compliant\" and Zero Data Retention."},"price_audio_in":{"value":null,"source":"https://docs.x.ai/docs/models","read":"2026-10-29","quote":"","note":"Not priced per audio token. See the price fact for the per-minute rate."},"price_audio_out":{"value":null,"source":"https://docs.x.ai/docs/models","read":"2026-10-29","quote":"","note":"Not priced per audio token. See the price fact for the per-minute rate."}},"terms":{"verdict":"restricts other use","source":"https://x.ai/legal/terms-of-service-enterprise","read":"2026-10-29","quote":"Scraping, harvesting or reselling any Input or Output, or distilling model data or Outputs"},"independent":[{"by":"tau-Voice (arXiv 2603.13686)","model":"grok-voice-agent","metric":"task completion pass@1 on realistic phone audio (noise, accents, interruptions), higher is better","value":"38%","source":"https://arxiv.org/abs/2603.13686","quote":"xAI 51% (-34, -40.1%) 38% (-46, -54.7%)","note":"Tested grok-voice-agent, an earlier version than this row's product. Table 6, all three domains (retail, airline, telecom), 8 kHz telephony with simulated callers. Higher is better.","read":"2026-10-30"},{"by":"tau-Voice (arXiv 2603.13686)","model":"grok-voice-agent","metric":"task completion pass@1 on clean phone audio, higher is better","value":"51%","source":"https://arxiv.org/abs/2603.13686","quote":"xAI 51% (-34, -40.1%) 38% (-46, -54.7%)","note":"Tested grok-voice-agent, an earlier version than this row's product. Table 6, all three domains (retail, airline, telecom), 8 kHz telephony with simulated callers. Higher is better.","read":"2026-10-30"}]},{"id":"kyutai-moshi","company":"Kyutai","product":"Moshi","layer":"s2s","homepage":"https://github.com/kyutai-labs/moshi","summary":"A free, open-weight research model that listens and speaks at once, you run it yourself on your own GPU.","open_weights":true,"facts":{"price":{"value":null,"source":"https://github.com/kyutai-labs/moshi","read":"2026-10-29","quote":"All models are released under the CC-BY 4.0 license.","note":"No hosted service or price. The weights are free to download under CC-BY 4.0, your cost is the GPU you run it on."},"tokens_per_minute":{"value":null,"source":"https://github.com/kyutai-labs/moshi","read":"2026-10-29","quote":"","note":"Not applicable to a self-hosted model."},"latency":{"value":"160 ms theoretical, about 200 ms in practice on an L4 GPU","source":"https://github.com/kyutai-labs/moshi","read":"2026-10-29","quote":"Moshi achieves a theoretical latency of 160ms (80ms for the frame size of Mimi + 80ms of acoustic delay), with a practical overall latency as low as 200ms on an L4 GPU.","note":"Vendor claim for the model alone on a GPU. It does not include phone network delay."},"en_gb":{"value":null,"source":"https://huggingface.co/kyutai/moshiko-pytorch-bf16","read":"2026-10-29","quote":"Language(s) (NLP): English","note":"English only. No accent or British English claim. The pre-training audio \"consists mostly of English speech\" and the model speaks with one voice per release."},"function_calling":{"value":null,"source":"https://github.com/kyutai-labs/moshi","read":"2026-10-29","quote":"","note":"The README does not mention tool or function calling."},"suitability":{"value":"Kyutai says research only","source":"https://huggingface.co/kyutai/moshiko-pytorch-bf16","read":"2026-10-29","quote":"This model is for research only and we do not recommend it for providing advices or to perform any professionnal duty.","note":""},"data_region":{"value":null,"source":"https://github.com/kyutai-labs/moshi","read":"2026-10-29","quote":"","note":"Self-hosted, so data stays wherever you run it. This is an inference from open weights, not a vendor statement."},"compliance":{"value":null,"source":"https://github.com/kyutai-labs/moshi","read":"2026-10-29","quote":"","note":"No compliance statement, open-source project."},"price_audio_in":{"value":null,"source":"https://github.com/kyutai-labs/moshi","read":"2026-10-29","quote":"","note":"Not priced per audio token. See the price fact for the per-minute rate."},"price_audio_out":{"value":null,"source":"https://github.com/kyutai-labs/moshi","read":"2026-10-29","quote":"","note":"Not priced per audio token. See the price fact for the per-minute rate."}},"independent":[{"by":"VoiceBench (arXiv 2410.17196)","metric":"overall score on spoken instructions, out of 100, higher is better","value":"27.47","source":"https://arxiv.org/abs/2410.17196","quote":"Moshi S. 2.01 1.60 15.01 16.27 24.04 26.15 6.38 13.76 44.23 27.47","note":"Table 3, speech-input row, Overall column. October 2024.","read":"2026-10-30"}]},{"id":"openai-gpt-6-luna","company":"OpenAI","product":"GPT-6 Luna","layer":"llm","homepage":"https://developers.openai.com/api/docs/models/gpt-6-luna","summary":"OpenAI's smallest, cheapest current text model, aimed at high-volume tasks.","open_weights":false,"facts":{"price_input":{"value":"$0.10 per 1M input tokens","amount":0.1,"currency":"USD","per":"1m_tokens_input","tier":"standard, prompts up to 272K tokens","source":"https://developers.openai.com/api/docs/models/gpt-6-luna","read":"2026-10-29","quote":"Text tokens Per 1M tokens Input $0.10 Cached input $0.01 Cache writes $0.125 Output $0.50","note":"Prompts over 272K input tokens cost 2x input and 1.5x output for the whole request. Cached input $0.01."},"price_output":{"value":"$0.50 per 1M output tokens","amount":0.5,"currency":"USD","per":"1m_tokens_output","tier":"standard, prompts up to 272K tokens","source":"https://developers.openai.com/api/docs/models/gpt-6-luna","read":"2026-10-29","quote":"Text tokens Per 1M tokens Input $0.10 Cached input $0.01 Cache writes $0.125 Output $0.50","note":""},"context_window":{"value":"1,050,000 tokens","source":"https://developers.openai.com/api/docs/models/gpt-6-luna","read":"2026-10-29","quote":"1,050,000 context window","note":""},"latency":{"value":null,"source":"https://developers.openai.com/api/docs/models/gpt-6-luna","read":"2026-10-29","quote":"","note":"No latency figure on the model or pricing pages."}},"terms":{"verdict":"restricts other use","source":"https://web.archive.org/web/20261028230302/https://openai.com/policies/services-agreement/","read":"2026-10-29","quote":"You may use OpenAI models for research, software development, model evaluation and benchmarking, and product development, within these boundaries."}},{"id":"google-gemini-3-5-flash-lite","company":"Google","product":"Gemini 3.5 Flash-Lite","layer":"llm","homepage":"https://ai.google.dev/gemini-api/docs/models/gemini-3.5-flash-lite","summary":"Google's fast, low-cost text model for high-volume work.","open_weights":false,"facts":{"price_input":{"value":"$0.30 per 1M input tokens","amount":0.3,"currency":"USD","per":"1m_tokens_input","tier":"paid tier, standard","source":"https://ai.google.dev/gemini-api/docs/pricing?hl=en","read":"2026-10-29","quote":"Gemini 3.5 Flash-Lite gemini-3.5-flash-lite Try it in Google AI Studio A cost-efficient model, optimized for high-volume agentic tasks, translation, and simple data processing. Standard Free Tier Paid Tier, per 1M tokens in USD Input price Free of charge $0.30 (text / image / video / audio)","note":"Same price for text, image, video and audio input. Batch is half."},"price_output":{"value":"$2.50 per 1M output tokens (thinking tokens included)","amount":2.5,"currency":"USD","per":"1m_tokens_output","tier":"paid tier, standard","source":"https://ai.google.dev/gemini-api/docs/pricing?hl=en","read":"2026-10-29","quote":"Output price (including thinking tokens) Free of charge $2.50 Context caching price Not available $0.03","note":""},"context_window":{"value":"1,048,576 input tokens","source":"https://ai.google.dev/gemini-api/docs/models/gemini-3.5-flash-lite?hl=en","read":"2026-10-29","quote":"Input token limit 1,048,576","note":"Output limit 65,536."},"latency":{"value":null,"source":"https://ai.google.dev/gemini-api/docs/models/gemini-3.5-flash-lite?hl=en","read":"2026-10-29","quote":"","note":"No latency figure on the model or pricing pages. Model page calls it low-latency."}},"terms":{"verdict":"restricts other use","source":"https://ai.google.dev/gemini-api/terms","read":"2026-10-29","quote":"Customer may itself (but may not permit a third party to): (a) conduct benchmark tests of the Services (each a “Test”); and (b) publicly disclose the results of such Tests only if (i) the public disclosure includes all necessary information to replicate the Tests, and (ii) Customer allows Google to conduct benchmark tests of Customer's publicly available products or services and publicly disclose the results of such tests."}},{"id":"anthropic-claude-haiku-5-5","company":"Anthropic","product":"Claude Haiku 5.5","layer":"llm","homepage":"https://platform.claude.com/docs/en/models/haiku-5-5/overview","summary":"Anthropic's smallest and fastest current model.","open_weights":false,"facts":{"price_input":{"value":"$0.10 per 1M input tokens (prompts up to 100,000 tokens)","amount":0.1,"currency":"USD","per":"1m_tokens_input","tier":"Claude API, global routing, prompts up to 100,000 tokens","source":"https://platform.claude.com/docs/en/about-claude/pricing","read":"2026-10-29","quote":"$0.10 / MTok for prompts up to 100,000 tokens $0.50 / MTok $0.125 / MTok $0.20 / MTok $0.01 / MTok $0.50 / MTok for prompts over 100,000 tokens $2.50 / MTok","note":"Row of the pricing table for Claude Haiku 5.5, columns are base input, cache writes (5 minute and 1 hour), cache hits and output. Over 100,000 prompt tokens the price is $0.50 input and $2.50 output. US-only inference adds 1.1x."},"price_output":{"value":"$0.50 per 1M output tokens (prompts up to 100,000 tokens)","amount":0.5,"currency":"USD","per":"1m_tokens_output","tier":"Claude API, global routing, prompts up to 100,000 tokens","source":"https://platform.claude.com/docs/en/about-claude/models/overview","read":"2026-10-29","quote":"From $0.10 / input MTok From $0.50 / output MTok","note":"The models overview prints the starting prices, the pricing page row for prompts up to 100,000 tokens ends with the same $0.50 output figure."},"context_window":{"value":"1M tokens","source":"https://platform.claude.com/docs/en/about-claude/models/overview","read":"2026-10-29","quote":"Context window 1M tokens 1M tokens 1M tokens 1M tokens","note":"Row of the comparison table, Haiku 5.5 is the last of four columns. The pricing page notes that Haiku 5.5 is the one model where prompts over 100,000 tokens cost more."},"latency":{"value":"Anthropic ranks it \"Fastest\" of its four current models","source":"https://platform.claude.com/docs/en/about-claude/models/overview","read":"2026-10-29","quote":"Comparative latency Slower Moderate Fast Fastest","note":"A relative label, not a time."}},"terms":{"verdict":"restricts other use","source":"https://www.anthropic.com/legal/commercial-terms","read":"2026-10-29","quote":"All consumer-facing chatbots or AI agents that communicate directly with external users must disclose to users that they are interacting with AI rather than a human."}},{"id":"groq-llama-3-1-8b-instant","company":"Groq (Meta Llama)","product":"Llama 3.1 8B Instant on GroqCloud","layer":"llm","homepage":"https://console.groq.com/docs/models","summary":"Meta's small Llama model on Groq's fast chips. Groq's pricing page does not print a pay-as-you-go price for its Llama models.","open_weights":true,"facts":{"price_input":{"value":null,"source":"https://console.groq.com/docs/models","read":"2026-10-29","quote":"Llama 3.1 8B Enterprise llama-3.1-8b-instant 560 Contact Sales Contact Sales 131,072 131,072","note":"The models page shows \"Contact Sales\" for price and rate limits on both Llama 3.1 8B and Llama 3.3 70B, tagged Enterprise. No list price is published for either Llama model. Groq's separate groq.com/pricing address redirects to the home page."},"price_output":{"value":null,"source":"https://console.groq.com/docs/models","read":"2026-10-29","quote":"Llama 3.1 8B Enterprise llama-3.1-8b-instant 560 Contact Sales Contact Sales 131,072 131,072","note":"Same as input: contact sales."},"context_window":{"value":"131,072 tokens","source":"https://console.groq.com/docs/models","read":"2026-10-29","quote":"Llama 3.1 8B Enterprise llama-3.1-8b-instant 560 Contact Sales Contact Sales 131,072 131,072","note":""},"speed":{"value":"560 tokens per second (vendor figure)","source":"https://console.groq.com/docs/models","read":"2026-10-29","quote":"Llama 3.1 8B Enterprise llama-3.1-8b-instant 560","note":"Output speed, not time to first token."}},"independent":[{"by":"Berkeley Function Calling Leaderboard (BFCL v4)","model":"Llama 3.1 8B Instruct on BFCL’s hosting","metric":"overall function-calling accuracy, higher is better","value":"25.83%","source":"https://gorilla.cs.berkeley.edu/data_overall.csv","quote":"85,25.83%,Llama-3.1-8B-Instruct (Prompt)","note":"Rank 85 of 109. Same weights, run in prompt mode by BFCL, not on Groq.","read":"2026-10-30"}]},{"id":"groq-gpt-oss-120b","company":"Groq","product":"GPT OSS 120B on GroqCloud","layer":"llm","homepage":"https://console.groq.com/docs/models","summary":"OpenAI's open-weight 120B model on Groq, included as a Groq text model with a printed price.","open_weights":true,"facts":{"price_input":{"value":"$0.15 per 1M input tokens","amount":0.15,"currency":"USD","per":"1m_tokens_input","tier":"developer plan, production model","source":"https://console.groq.com/docs/models","read":"2026-10-29","quote":"GPT OSS 120B openai/gpt-oss-120b 500 $0.15 input $0.60 output 250K TPM 1K RPM 131,072 65,536","note":""},"price_output":{"value":"$0.60 per 1M output tokens","amount":0.6,"currency":"USD","per":"1m_tokens_output","tier":"developer plan, production model","source":"https://console.groq.com/docs/models","read":"2026-10-29","quote":"GPT OSS 120B openai/gpt-oss-120b 500 $0.15 input $0.60 output 250K TPM 1K RPM 131,072 65,536","note":""},"context_window":{"value":"131,072 tokens","source":"https://console.groq.com/docs/models","read":"2026-10-29","quote":"GPT OSS 120B openai/gpt-oss-120b 500 $0.15 input $0.60 output 250K TPM 1K RPM 131,072 65,536","note":""},"speed":{"value":"500 tokens per second (vendor figure)","source":"https://console.groq.com/docs/models","read":"2026-10-29","quote":"GPT OSS 120B openai/gpt-oss-120b 500","note":"Output speed, not time to first token."}}},{"id":"cerebras-gpt-oss-120b","company":"Cerebras","product":"GPT OSS 120B on Cerebras Inference","layer":"llm","homepage":"https://www.cerebras.ai/pricing","summary":"OpenAI's open-weight 120B model on Cerebras chips. Cerebras's self-serve pricing lists no Llama model.","open_weights":true,"facts":{"price_input":{"value":"$0.35 per 1M input tokens","amount":0.35,"currency":"USD","per":"1m_tokens_input","tier":"developer tier, pay as you go","source":"https://www.cerebras.ai/pricing","read":"2026-10-29","quote":"GPT OSS 120B ~3000 tokens/s $0.35/M tokens $0.75/M tokens","note":"The pricing page lists only GPT OSS 120b and Qwen 3.8 27B on the developer tier, so there is no Llama price to report. Read in a real browser (the table loads by script)."},"price_output":{"value":"$0.75 per 1M output tokens","amount":0.75,"currency":"USD","per":"1m_tokens_output","tier":"developer tier, pay as you go","source":"https://www.cerebras.ai/pricing","read":"2026-10-29","quote":"GPT OSS 120B ~3000 tokens/s $0.35/M tokens $0.75/M tokens","note":""},"context_window":{"value":"65k tokens on the free tier, 131k on paid tiers","source":"https://inference-docs.cerebras.ai/models/overview","read":"2026-10-29","quote":"OpenAI GPT OSS gpt-oss-120b 120 billion 65k / 131k ~3000","note":""},"speed":{"value":"about 3,000 tokens per second (vendor figure)","source":"https://www.cerebras.ai/pricing","read":"2026-10-29","quote":"GPT OSS 120B ~3000 tokens/s","note":"Output speed, not time to first token."}}},{"id":"mistral-small-4","company":"Mistral AI","product":"Mistral Small 4","layer":"llm","homepage":"https://docs.mistral.ai/models/mistral-small-4-0-26-03","summary":"Mistral's cost-focused model: a large model that only uses a small part of itself per token.","open_weights":true,"facts":{"price_input":{"value":"$0.15 per 1M input tokens","amount":0.15,"currency":"USD","per":"1m_tokens_input","tier":"API list price","source":"https://docs.mistral.ai/models/mistral-small-4-0-26-03","read":"2026-10-29","quote":"Context i 256k Price i USD EUR $0.15 /M Tokens $0.6 /M Tokens","note":"The page also shows EUR prices behind a toggle, the USD tab was read. The two figures are input then output."},"price_output":{"value":"$0.60 per 1M output tokens","amount":0.6,"currency":"USD","per":"1m_tokens_output","tier":"API list price","source":"https://docs.mistral.ai/models/mistral-small-4-0-26-03","read":"2026-10-29","quote":"Context i 256k Price i USD EUR $0.15 /M Tokens $0.6 /M Tokens","note":""},"context_window":{"value":"256k tokens","source":"https://docs.mistral.ai/models/mistral-small-4-0-26-03","read":"2026-10-29","quote":"Context i 256k Price i USD","note":""},"latency":{"value":null,"source":"https://docs.mistral.ai/models/mistral-small-4-0-26-03","read":"2026-10-29","quote":"","note":"No latency figure on the model page."}}},{"id":"deepseek-v4-1-flash","company":"DeepSeek","product":"DeepSeek-V4.1-Flash-1017 (API name deepseek-flash)","layer":"llm","homepage":"https://api-docs.deepseek.com/quick_start/pricing","summary":"DeepSeek's low-priced, fast model, charged at half rate outside its peak hours (01:00-04:00 and 06:00-10:00 UTC, Monday to Friday).","open_weights":false,"facts":{"price_input":{"value":"$0.30 per 1M input tokens at peak, $0.15 off-peak (cache miss)","amount":0.3,"currency":"USD","per":"1m_tokens_input","tier":"peak rate, cache miss; off-peak 0.15","source":"https://api-docs.deepseek.com/quick_start/pricing/","read":"2026-10-29","quote":"1M INPUT TOKENS (CACHE MISS) OFF-PEAK $0.15 $0.66 PEAK $0.3 $1.32","note":"The first figure in each pair is deepseek-flash, the second deepseek-v4-pro. Cache hit is $0.006 peak, $0.003 off-peak. Peak hours are 01:00-04:00 and 06:00-10:00 UTC, Monday to Friday, excluding Chinese public holidays, all other hours are off-peak."},"price_output":{"value":"$1.20 per 1M output tokens at peak, $0.60 off-peak","amount":1.2,"currency":"USD","per":"1m_tokens_output","tier":"peak rate; off-peak 0.6","source":"https://api-docs.deepseek.com/quick_start/pricing/","read":"2026-10-29","quote":"1M OUTPUT TOKENS OFF-PEAK $0.6 $1.98 PEAK $1.2 $3.96","note":""},"context_window":{"value":"1M tokens","source":"https://api-docs.deepseek.com/quick_start/pricing/","read":"2026-10-29","quote":"CONTEXT LENGTH 1M","note":"Max output 384K."},"latency":{"value":null,"source":"https://api-docs.deepseek.com/quick_start/pricing/","read":"2026-10-29","quote":"","note":"No latency figure on the pricing page."}}},{"id":"twilio-voice","company":"Twilio","product":"Programmable Voice and Elastic SIP Trunking (UK numbers)","layer":"carrier","homepage":"https://www.twilio.com/","summary":"Large US phone-API company that rents UK numbers and carries calls over its API or over SIP trunks.","facts":{"uk_number_price":{"value":"$3.50 per month","amount":3.5,"currency":"USD","per":"number_month","tier":"UK local number, pay as you go","source":"https://www.twilio.com/en-us/voice/pricing/gb","read":"2026-10-29","quote":"Local numbers $3.50 / mo","note":"Page header reads 'Pricing current as of August 2026'. UK mobile numbers $2.50 / mo, toll-free $2.70 / mo."},"uk_inbound":{"value":"$0.0100 per minute","amount":0.01,"currency":"USD","per":"call_minute","tier":"receiving on a UK local number","source":"https://www.twilio.com/en-us/voice/pricing/gb","read":"2026-10-29","quote":"Local numbers $0.0100 / min + $3.50 / mo","note":""},"uk_outbound_landline":{"value":"$0.0158 per minute","amount":0.0158,"currency":"USD","per":"call_minute","tier":"calls to UK landlines (the 'Local calls' row)","source":"https://www.twilio.com/en-us/voice/pricing/gb","read":"2026-10-29","quote":"Local calls $0.0158 / min","note":"Calls to UK landlines from a Twilio-hosted origin. Separate rows apply from 'Surcharged Zone' origins: 'United Kingdom - From Surcharged Zone 1 $0.2293 / min'."},"uk_outbound_mobile":{"value":"$0.0305 per minute","amount":0.0305,"currency":"USD","per":"call_minute","tier":"calls to UK mobiles (the 'Mobile calls' row)","source":"https://www.twilio.com/en-us/voice/pricing/gb","read":"2026-10-29","quote":"Mobile calls $0.0305 / min","note":"'United Kingdom - Mobile - From Surcharged Zone 1 $0.4730 / min' and 'Mobile - Other $0.3200 / min' also listed."},"sip_trunking":{"value":"yes","source":"https://www.twilio.com/en-us/sip-trunking/pricing/gb","read":"2026-10-29","quote":"Elastic SIP Trunking costs are tied to minutes, call paths, and geographical regions","note":"UK termination on Elastic SIP Trunking: 'Starting at $0.0118 / min'. The voice pricing page also lists BYOC trunking at $0.0040 / min each way."},"uk_number_rules":{"value":"UK address required (a valid United Kingdom address), individuals give government ID plus proof of address such as a utility bill, tax notice, rent receipt or title deed, businesses give registration details and a UK business address.","source":"https://www.twilio.com/en-us/guidelines/gb/regulatory","read":"2026-10-29","quote":"Proof of Address Government-issued ID showing local address Utility bill Tax notice Rent receipt Title deed","note":"Same page, businesses: 'Business Address Must be a valid United Kingdom address.' Applies to UK local and national numbers, mobile and toll-free have a worldwide-address rule."},"data_region":{"value":"Ireland (IE1, Dublin) region offers Programmable Voice and Elastic SIP Trunking, but the US1 region is the default unless you choose one","source":"https://www.twilio.com/docs/global-infrastructure/regional-product-and-feature-availability","read":"2026-10-29","quote":"Ireland (IE1) Region This Region is located in Dublin.","note":"On https://www.twilio.com/docs/global-infrastructure/understanding-twilio-regions: 'US1 is the default Region where all workloads are processed if a Region isn't explicitly selected.' That page also warns Twilio 'doesn't guarantee that all data will remain within your selected Region' during the initial rollout, and some voice features are unsupported in IE1."},"compliance":{"value":"HIPAA-eligible products, SOC 2 Type II, PCI DSS Level 1, ISO/IEC 27001, 27017 and 27018, Binding Corporate Rules","source":"https://www.twilio.com/en-us/security","read":"2026-10-29","quote":"Please note the credentials listed do not apply to every product across Twilio, Segment, and SendGrid.","note":"Credentials listed on the page: HIPAA eligible products and services, SOC 2 Type II, PCI DSS Level 1, Binding Corporate Rules, ISO/IEC 27017:2015, 27018:2019 and 27001:2013."}},"terms":{"verdict":"no clause found","source":"https://www.twilio.com/en-us/legal/tos","read":"2026-10-29","quote":"Reverse Engineering. Reverse engineering, copying, disassembling, or decompiling the Services."}},{"id":"telnyx-voice","company":"Telnyx","product":"Voice API and Elastic SIP Trunking (UK numbers)","layer":"carrier","homepage":"https://telnyx.com/","summary":"Phone carrier with its own network that rents UK numbers and carries calls over SIP trunks or its call API.","facts":{"uk_number_price":{"value":"from $1 per month","amount":1,"currency":"USD","per":"number_month","tier":"UK local number, 'start at' price","source":"https://telnyx.com/phone-numbers/united-kingdom","read":"2026-10-29","quote":"Local numbers for United Kingdom start at $1 per month.","note":"No exact UK-only table found, the global numbers page (https://telnyx.com/pricing/numbers) also says 'From $1 per month*'. Currency not named on the page, shown with a $ sign (the rates table has a USD selector)."},"uk_inbound":{"value":"starting at $0.005 per minute","amount":0.005,"currency":"USD","per":"call_minute","tier":"local calls received on a UK number, pay as you go, 'starting at'","source":"https://telnyx.com/pricing/elastic-sip","read":"2026-10-29","quote":"Local calls Starting at $0.005 per minute","note":"Read on the SIP Trunking pricing page with the page's country box set to United Kingdom (the heading changes to 'Pay as you go for United Kingdom'). The page loads one country at a time with script, the default view is the United States. Inbound can instead be billed per channel ($12 per month for the first 10 channels). The Voice API adds $0.002 per minute on top of the SIP fee."},"uk_outbound_landline":{"value":"$0.0042 per minute","amount":0.0042,"currency":"USD","per":"call_minute","tier":"outbound to UK fixed lines, pay as you go","source":"https://telnyx.com/pricing/elastic-sip","read":"2026-10-29","quote":"Local - Fixed $0.0042 per minute","note":"Same UK view of the SIP Trunking pricing page. Volume tiers and carrier pass-through may change it ('Rates may vary by destination and volume tier')."},"uk_outbound_mobile":{"value":"$0.02 per minute","amount":0.02,"currency":"USD","per":"call_minute","tier":"outbound to UK mobiles, pay as you go","source":"https://telnyx.com/pricing/elastic-sip","read":"2026-10-29","quote":"Local - Mobile $0.02 per minute","note":"Same UK view of the SIP Trunking pricing page."},"sip_trunking":{"value":"yes","source":"https://telnyx.com/pricing/voice-api.md","read":"2026-10-29","quote":"Elastic SIP trunks on the Telnyx network.","note":"Secure trunking is listed as Free on the SIP Trunking pricing page."},"uk_number_rules":{"value":"Identity and address documents are needed: a passport or ID copy (or company registration certificate for a business), plus a worldwide-format address and a utility bill less than 3 months old.","source":"https://support.telnyx.com/en/articles/1311457-united-kingdom-uk-did-requirements","read":"2026-10-29","quote":"In order to purchase an United Kingdom number you will need to provide the following:","note":"Under 'Local Numbers in the United Kingdom': 'Worldwide Address (street, building number, postal code, city, and country)' and 'A copy of a utility bill (less than 3 months old)'. Other sources have shown a stricter UK-address rule for this page, so re-check before relying on it."},"data_region":{"value":"EU endpoint processes in Frankfurt, storage in the EU needs a separate Data Locality setting. London is not named as a processing site.","source":"https://developers.telnyx.com/docs/account-setup/regional-endpoints.md","read":"2026-10-29","quote":"the EU PoP runs in Frankfurt, Germany.","note":"Same page: 'Sending to `api.telnyx.eu` alone does **not** change where data is stored at rest' and processing location 'is not contractually guaranteed'. A search result also showed London, Frankfurt and Amsterdam as SIP media sites on a separate IP-allowlist page, which We did not open."},"compliance":{"value":"ISO, PCI, HIPAA, GDPR, SOC 2 Type II (as printed in the site footer)","source":"https://telnyx.com/data-privacy","read":"2026-10-29","quote":"ISO • PCI • HIPAA • GDPR • SOC2 Type II","note":"Footer line. The same page says Telnyx follows GDPR and offers contract terms for customers' GDPR duties."}},"terms":{"verdict":"restricts other use","source":"https://telnyx.com/terms-and-conditions-of-service","read":"2026-10-29","quote":"auto-dialing or predictive-dialing (sometimes referred to as “robo-dialing”) to place unsolicited calls, in violation of any Laws, or in a manner that results in Abandoned Calls (as defined below), dead air, complaints or other calling patterns that Telnyx in its sole discretion deems abusive"}},{"id":"vonage-voice","company":"Vonage (Communications APIs)","product":"Voice API, Programmable SIP and virtual numbers","layer":"carrier","homepage":"https://www.vonage.com/communications-apis/","summary":"Phone-API company that rents UK numbers and carries calls through its Voice API and SIP.","facts":{"uk_number_price":{"value":"$1.22 per month","amount":1.22,"currency":"USD","per":"number_month","tier":"local number, mobile and landline, voice","source":"https://www.vonage.com/communications-apis/voice/pricing/","read":"2026-10-29","quote":"LOCAL NUMBER - MOBILE & LANDLINE Voice $1.22 per month","note":"Read after typing United Kingdom into the page's Country/Region box (page heading: 'Finding the best prices in United Kingdom'). The card also shows 'Initial set up fee $1.22 ... Setup fee may apply'. The page blocks plain scripted fetches, so this was read in a browser, and the default view without a country is the United States."},"uk_inbound":{"value":"from $0.00872 per minute","amount":0.00872,"currency":"USD","per":"call_minute","tier":"'Receive a call from' on the International calling card, UK selected","source":"https://www.vonage.com/communications-apis/voice/pricing/","read":"2026-10-29","quote":"INTERNATIONAL CALLING (PSTN) ... Receive a call from $0.00872 per minute","note":"Open rate. A separate Domestic calling card lists 'Receive a call $0.00480 per minute' but says 'To get access to these rates, please contact sales.' (see uk_inbound_domestic)."},"uk_outbound_landline":{"value":"from $0.01454 per minute","amount":0.01454,"currency":"USD","per":"call_minute","tier":"'Make a call to a landline number from' on the International calling card, UK selected","source":"https://www.vonage.com/communications-apis/voice/pricing/","read":"2026-10-29","quote":"Make a call to a landline number from $0.01454 per minute","note":"Same page adds: 'In response to regulatory changes impacting the pricing of international calls, it is necessary for Vonage to add a surcharge for some voice calls terminating in the UK.' Domestic rate in uk_outbound_landline_domestic."},"uk_outbound_mobile":{"value":"from $0.08719 per minute","amount":0.08719,"currency":"USD","per":"call_minute","tier":"'Make a call to a mobile number from' on the International calling card, UK selected","source":"https://www.vonage.com/communications-apis/voice/pricing/","read":"2026-10-29","quote":"Make a call to a mobile number from $0.08719 per minute","note":"Domestic rate in uk_outbound_mobile_domestic."},"uk_inbound_domestic":{"value":"$0.00480 per minute","amount":0.0048,"currency":"USD","per":"call_minute","tier":"Domestic calling card, UK selected, gated","source":"https://www.vonage.com/communications-apis/voice/pricing/","read":"2026-10-29","quote":"Receive a call $0.00480 per minute To get access to these rates, please contact sales.","note":"Printed on the page but the page says access needs a sales contact, so it is not a self-serve list price."},"uk_outbound_landline_domestic":{"value":"$0.00970 per minute","amount":0.0097,"currency":"USD","per":"call_minute","tier":"Domestic calling card, UK selected, gated","source":"https://www.vonage.com/communications-apis/voice/pricing/","read":"2026-10-29","quote":"Make a call to a landline number $0.00970 per minute","note":"Gated: 'To get access to these rates, please contact sales.'"},"uk_outbound_mobile_domestic":{"value":"$0.02676 per minute","amount":0.02676,"currency":"USD","per":"call_minute","tier":"Domestic calling card, UK selected, gated","source":"https://www.vonage.com/communications-apis/voice/pricing/","read":"2026-10-29","quote":"Make a call to a mobile number $0.02676 per minute","note":"Gated: 'To get access to these rates, please contact sales.'"},"sip_trunking":{"value":"yes (SIP endpoint calls at $0.00492 per minute)","source":"https://www.vonage.com/communications-apis/voice/pricing/","read":"2026-10-29","quote":"Make/receive a call to SIP endpoint $0.00492 per minute","note":"Listed under Internet calling. This is the Voice API's SIP connection, which carries the call to your own SIP system."},"uk_number_rules":{"value":"UK address required","source":"https://api.support.vonage.com/hc/en-us/articles/115011761808-Voice-Numbers-Features-Overview","read":"2026-10-29","quote":"United Kingdom Yes Landline Yes Dashboard/API Instant UK address required 50","note":"Row from Vonage's voice numbers table. Columns: international reach for inbound voice call, number type, address required, availability, order lead time, notes and purchase requirements, concurrent inbound calls per number. The UK-specific features page (api.support.vonage.com .../20052374957212) gives no address or ID rules."},"data_region":{"value":"EU endpoints process calls in Dublin, with Frankfurt as fallback, an application can be pinned to the EU region","source":"https://developer.vonage.com/en/voice/voice-api/concepts/regions","read":"2026-10-29","quote":"api-eu.vonage.com : Dublin, or Frankfurt in case of fallback","note":"Same page: 'user data from one geo-location is not accessible in another due to compliance requirements'. Default api.nexmo.com sets up call legs in Virginia, per a search summary of this page."},"compliance":{"value":"Core API (SMS/Voice) lists ISO 27001, SOC 2 Type II and a HIPAA attestation","source":"https://developer.vonage.com/en/getting-started/concepts/regulatory-certifications","read":"2026-10-29","quote":"This page documents the certifications held by Vonage for the following lines of business: Core API (SMS/Voice) & Vonage Network APIs (VNA), Vonage AI Media Analyzer, and Vonage AI Studio.","note":"Table rows under Core API (SMS/Voice): 'ISO 27001', 'SOC 2 Type II', 'HIPAA Attestation of Compliance'. The page notes caveats per product."}},"terms":{"verdict":"restricts publishing","source":"https://www.vonage.com/legal/communications-apis/terms-of-use/","read":"2026-10-29","quote":"Benchmarking, tampering with, unauthorized testing of, reverse-engineering, decompiling, or otherwise using the Services in order to discover limitations or vulnerabilities, or evade filtering capabilities."}},{"id":"plivo-voice","company":"Plivo","product":"Voice API and Zentrunk SIP trunking (UK numbers)","layer":"carrier","homepage":"https://www.plivo.com/","summary":"Phone-API company that rents UK numbers and carries calls over its API or SIP trunks.","facts":{"uk_number_price":{"value":"$0.85 per month","amount":0.85,"currency":"USD","per":"number_month","tier":"UK local number","source":"https://www.plivo.com/voice/pricing/gb/","read":"2026-10-29","quote":"Local Numbers $0.85/month","note":"Page footer: 'Rates as of May 2026, USD/min.' (older than the 29 October read date). UK mobile numbers $0.90/month, toll-free $1.00/month."},"uk_inbound":{"value":"$0.0055 per minute","amount":0.0055,"currency":"USD","per":"call_minute","tier":"receiving on a UK local number","source":"https://www.plivo.com/voice/pricing/gb/","read":"2026-10-29","quote":"Local Calls Starts at $0.0075/min View detailed network pricing $0.0055/min","note":"The $0.0055 is the 'To Receive Calls (Inbound)' column for local calls, $0.0075 is the outbound column."},"uk_outbound_landline":{"value":"starts at $0.0075 per minute","amount":0.0075,"currency":"USD","per":"call_minute","tier":"outbound, UK local (landline) calls, 'starts at'","source":"https://www.plivo.com/voice/pricing/gb/","read":"2026-10-29","quote":"Local Calls Starts at $0.0075/min","note":"Network table splits by origin and carrier, on the SIP trunking page the UK fixed rate is $0.0072/min."},"uk_outbound_mobile":{"value":"$0.0270 per minute","amount":0.027,"currency":"USD","per":"call_minute","tier":"outbound to UK mobiles","source":"https://www.plivo.com/voice/pricing/gb/","read":"2026-10-29","quote":"Mobile Calls $0.0270/min $0.0075/min","note":"First figure is outbound, second is inbound for mobile numbers. The network table lists higher prices for some UK mobile networks, for example 'United Kingdom - Mobile - Major Carriers ... $0.0270/min' for calls from the EEA."},"sip_trunking":{"value":"yes","source":"https://www.plivo.com/sip-trunking/pricing/gb/","read":"2026-10-29","quote":"Competitive pay-as-you-go SIP Trunking pricing with add-on features included.","note":"Zentrunk UK rates: local outbound 'Starts at $0.0072/min', inbound $0.0030/min, UK local number $0.85/month."},"uk_number_rules":{"value":"No compliance documents required for UK numbers, per Plivo's numbers docs","source":"https://www.plivo.com/docs/numbers","read":"2026-10-29","quote":"Countries not listed below (US, Canada, UK, Australia) do not require compliance documents.","note":"This differs from Twilio and Telnyx, which do ask UK buyers for ID and address proof. Reached through a redirect from support.plivo.com article 360041404752. The same docs page also says 'Mobile numbers support SMS and voice. Plivo offers mobile numbers in select countries only (UK, Australia).'"},"data_region":{"value":null,"source":"https://www.plivo.com/security/","read":"2026-10-29","quote":"","note":"No UK or EU voice region or media-routing statement on the security page or the UK pricing pages. Not found, not the same as 'none'."},"compliance":{"value":"GDPR-aligned, SOC 2 certified, HIPAA/HITECH with a Business Associate Agreement, EU-US Data Privacy Framework participant","source":"https://www.plivo.com/security/","read":"2026-10-29","quote":"Plivo is SOC 2 certified. We're audited annually by an independent auditor.","note":"Also on the page: 'Plivo systems are compliant with the data protection principles of the European Union's General Data Protection Regulation.' and 'Plivo participates in the EU–U.S. Data Privacy Framework (DPF).'"}},"terms":{"verdict":"restricts other use","source":"https://www.plivo.com/legal/tos/","read":"2026-10-29","quote":"(iii) access, search, or attempt to access or search the Plivo Cloud by any means (automated or otherwise) other than through our currently available, published interfaces that we provide (and only pursuant to those terms and conditions), unless you have been specifically allowed to do so in a separate agreement with us"}},{"id":"signalwire-voice","company":"SignalWire","product":"Voice and SIP (UK numbers by request)","layer":"carrier","homepage":"https://signalwire.com/","summary":"Phone and voice-AI platform built on FreeSWITCH, UK numbers can be ordered but no UK prices are printed.","facts":{"uk_number_price":{"value":null,"source":"https://signalwire.com/pricing/rates","read":"2026-10-29","quote":"Local (10DLC) $0.50","note":"Public pricing and the full rate deck cover United States and Canada only ($0.50 per month is a US local number). UK numbers come through a request form or sales."},"uk_inbound":{"value":null,"source":"https://signalwire.com/pricing/rates","read":"2026-10-29","quote":"PSTN (10DLC) — Inbound $0.0066 /min","note":"Printed rates are US numbers. No UK inbound rate on the rate deck page."},"uk_outbound_landline":{"value":null,"source":"https://signalwire.com/pricing/rates","read":"2026-10-29","quote":"PSTN (10DLC) — Outbound $0.0080 /min","note":"Printed rates are US. No UK landline rate published."},"uk_outbound_mobile":{"value":null,"source":"https://signalwire.com/pricing/rates","read":"2026-10-29","quote":"PSTN (10DLC) — Outbound $0.0080 /min","note":"No UK mobile rate published."},"sip_trunking":{"value":"yes (SIP / WebRTC transport at $0.0030 per minute, a US/Canada rate)","source":"https://signalwire.com/pricing/rates","read":"2026-10-29","quote":"SIP / WebRTC $0.0030 /min","note":"Voice transport rate on the rate deck page, which prices US and Canada."},"uk_number_rules":{"value":null,"source":"https://signalwire.com/docs/platform/how-to-enable-international-services.md","read":"2026-10-29","quote":"SignalWire officially supports acquiring numbers in the following countries:","note":"The docs list United Kingdom among supported countries and say 'Please submit a request form to procure an International number(s).' International services are off by default ('By default, SignalWire Spaces are only enabled for the US and Canada') and not enabled for trial spaces. No address or ID rules published."},"data_region":{"value":"Claims data residency across 30+ availability zones, no UK or London region named on the page","source":"https://signalwire.com/products/baa-hipaa-compliance","read":"2026-10-29","quote":"Data residency across 30+ availability zones","note":"Marketing line only, no list of regions."},"compliance":{"value":"SOC 2 Type II, HIPAA BAA, GDPR, PCI, ISO 27001","source":"https://signalwire.com/products/baa-hipaa-compliance","read":"2026-10-29","quote":"SignalWire maintains SOC 2 Type II attestation and ISO/IEC 27001:2022 certification, supports GDPR and PCI requirements, and offers HIPAA Business Associate Agreements (BAAs) for healthcare and other regulated customers.","note":""}},"terms":{"verdict":"restricts publishing","source":"https://info.signalwire.com/hubfs/PDFs/SignalWire-Cloud-Agreement.pdf","read":"2026-10-29","quote":"Customer agrees that all information regarding its use of Beta Services, including Customer’s experience with and opinions of Beta Services, will be deemed SignalWire’s confidential information, subject to protection under Section 8. Customer shall not disclose such testing results or experiences with any third party or use them for any purpose other than providing feedback to SignalWire."}},{"id":"sinch-voice","company":"Sinch","product":"Voice API (UK numbers)","layer":"carrier","homepage":"https://sinch.com/","summary":"Swedish phone-API company that rents UK numbers and carries calls through its Voice API.","facts":{"uk_number_price":{"value":"$0.8 per month","amount":0.8,"currency":"USD","per":"number_month","tier":"local business number, UK selected","source":"https://sinch.com/pricing/voice/","read":"2026-10-29","quote":"Monthly Fee $0.8 per month","note":"Read after choosing United Kingdom in the page's Country box (the page redirects to https://sinch.com/voice/pricing/). Also shows 'Setup Fee $0.8 One time fee'. The default view without a country shows $0.00 placeholders. Currency box reads USD."},"uk_inbound":{"value":"$0.008 per minute","amount":0.008,"currency":"USD","per":"call_minute","tier":"Domestic calling, receive calls, UK selected","source":"https://sinch.com/pricing/voice/","read":"2026-10-29","quote":"Receive calls $0.008 Per minute","note":""},"uk_outbound_landline":{"value":null,"source":"https://sinch.com/pricing/voice/","read":"2026-10-29","quote":"Make calls (prices from) $0.01 Per minute","note":"Sinch prints one outbound figure 'from $0.01 per minute' for domestic calling, not split into landline and mobile. A rate card is behind 'Download Voice API pricing'."},"uk_outbound_mobile":{"value":null,"source":"https://sinch.com/pricing/voice/","read":"2026-10-29","quote":"Make calls (prices from) $0.01 Per minute","note":"No separate UK mobile figure on the page."},"uk_outbound_from":{"value":"from $0.01 per minute","amount":0.01,"currency":"USD","per":"call_minute","tier":"domestic calling, 'prices from', UK selected","source":"https://sinch.com/pricing/voice/","read":"2026-10-29","quote":"Make calls (prices from) $0.01 Per minute","note":"One figure for domestic outbound, not split by landline or mobile."},"sip_trunking":{"value":"Voice API SIP interface yes, Elastic SIP Trunking only in the US and Canada","source":"https://sinch.com/pricing/voice/","read":"2026-10-29","quote":"Our Elastic SIP Trunking service can connect your existing PBX or contact center to our global network. Elastic SIP Trunk is available in US and Canada.","note":"The Voice API 'SIP interface' (UK selected) is priced 'Make calls $0.0025 Per minute' and 'Receive calls $0.0025 Per minute'."},"uk_number_rules":{"value":null,"source":"https://sinch.com/pricing/voice/","read":"2026-10-29","quote":"Numbers Get a local business number to connect with your customers.","note":"No UK address or ID rule on the pricing page. A Sinch support article says number cost depends on capability, type and country but gives no UK document rules."},"data_region":{"value":"Voice API has a Europe 1 (Central) region, the page does not say where in Europe or whether media stays there","source":"https://developers.sinch.com/docs/voice-2.0/api-reference/voice/section/core-concepts/regions","read":"2026-10-29","quote":"Europe 1 - Central https://eu1.voice.api.sinch.com","note":"Global endpoint 'Redirected by Sinch to the closest region'. No UK region listed."},"compliance":{"value":"ISO 27001, SOC 2 Type 1, GDPR, CCPA (voice product page)","source":"https://sinch.com/voice/security/","read":"2026-10-29","quote":"Sinch holds ISO 27001 and SOC 2 Type 1 certifications and meets regulations like GDPR, CCPA, DORA, Kari’s Law, and RAY BAUM’S Act.","note":"Another line on the same page: 'Yes. Sinch Voice is GDPR and CCPA compliant and ISO 27001 certified.' Other Sinch pages mention SOC 2 Type II for messaging, the voice page says Type 1."}},"terms":{"verdict":"restricts other use","source":"https://sinch.com/legal/terms-and-conditions/other-sinch-terms-conditions/general-terms-and-conditions/","read":"2026-10-29","quote":"Customer will not, directly or indirectly, use the AI Components: (i) to reverse engineer, decompile, discover or extract any underlying components of the models, algorithms, and systems; (ii) to develop its own artificial intelligence products, services, features, or technology, or any machine learning models or technology; (iii) to mislead anyone that an Output was human generated; or (iv) in any manner that violates applicable laws and regulations."}},{"id":"bandwidth-voice","company":"Bandwidth","product":"Voice and Global SIP Trunking","layer":"carrier","homepage":"https://www.bandwidth.com/","summary":"US phone carrier that lists the UK in its international coverage but publishes no UK prices.","facts":{"uk_number_price":{"value":null,"source":"https://www.bandwidth.com/pricing/","read":"2026-10-29","quote":"Volume based pricing","note":"No public per-country price list. The coverage page (https://www.bandwidth.com/coverage/) lists United Kingdom among countries with local, national and mobile number columns, but the filter output did not show UK rows."},"uk_inbound":{"value":null,"source":"https://www.bandwidth.com/pricing/","read":"2026-10-29","quote":"Volume based pricing","note":"No UK inbound rate printed."},"uk_outbound_landline":{"value":null,"source":"https://www.bandwidth.com/pricing/","read":"2026-10-29","quote":"Volume based pricing","note":"No UK outbound rate printed."},"uk_outbound_mobile":{"value":null,"source":"https://www.bandwidth.com/pricing/","read":"2026-10-29","quote":"Volume based pricing","note":"No UK mobile rate printed."},"sip_trunking":{"value":"yes (Global SIP Trunking listed as a product)","source":"https://www.bandwidth.com/pricing/","read":"2026-10-29","quote":"Global SIP Trunking","note":"Navigation entry. The coverage page title reads 'Global Network Coverage in 65+ countries | 40+ PSTN replacement countries', whether the UK is in the 40+ PSTN-replacement set was not stated."},"uk_number_rules":{"value":null,"source":"https://www.bandwidth.com/coverage/","read":"2026-10-29","quote":"United Kingdom","note":"No UK number-issue rules found on the pages I could read."},"data_region":{"value":null,"source":"https://www.bandwidth.com/coverage/","read":"2026-10-29","quote":"","note":"No UK or EU media-region statement found."},"compliance":{"value":null,"source":"https://www.bandwidth.com/pricing/","read":"2026-10-29","quote":"","note":"Not checked on a dedicated trust page, the pricing page carries no compliance claim."}}},{"id":"simwood-carrier","company":"Simwood","product":"UK carrier services and SIP trunks","layer":"carrier","homepage":"https://simwood.com/","summary":"Bristol-based UK carrier, carrier rates are shown only to account holders, and its public price list covers a separate hosted phone service that includes SIP trunks.","facts":{"uk_number_price":{"value":null,"source":"https://simwood.com/pricing","read":"2026-10-29","quote":"Please contact our sales team if you are not a customer and would like to see our rates.","note":"Carrier-services rates are behind the portal. Public Hosted partner list (https://cdn.simwood.com/docs/hosted_pricing.pdf) shows 'Additional DDI 1.50' per month in GBP, but that is a hosted-PBX partner plan, not the carrier price a developer would pay, so it is not used as a value."},"uk_inbound":{"value":null,"source":"https://simwood.com/pricing","read":"2026-10-29","quote":"Full rate sheets with codes are now only available through the portal or from our API for active customers.","note":"No public per-minute rate."},"uk_outbound_landline":{"value":null,"source":"https://simwood.com/pricing","read":"2026-10-29","quote":"Full rate sheets with codes are now only available through the portal or from our API for active customers.","note":"No public per-minute rate."},"uk_outbound_mobile":{"value":null,"source":"https://simwood.com/pricing","read":"2026-10-29","quote":"Full rate sheets with codes are now only available through the portal or from our API for active customers.","note":"No public per-minute rate."},"sip_trunking":{"value":"SIP trunks are sold on Simwood Hosted (a separate service), not confirmed for Simwood's carrier services on the public pages","source":"https://cdn.simwood.com/docs/hosted_pricing.pdf","read":"2026-10-29","quote":"SIP Trunk (per channel) 2.00 5.00 4.00","note":"Simwood Hosted partner price list (dated 26 July 2025), GBP per channel per month across Pearl, Emerald, Ruby plans (Diamond is POA). The PDF says its charges apply to Simwood Hosted customers and that 'Simwood Carrier Services customers are subject to their own pricing and terms'. The simwood.com home page lists SIP trunks only under Hosted."},"uk_number_rules":{"value":null,"source":"https://simwood.com/pricing","read":"2026-10-29","quote":"","note":"No Ofcom address or ID rule stated on the rates page."},"data_region":{"value":null,"source":"https://simwood.com/pricing","read":"2026-10-29","quote":"Simwood eSMS Limited Simwood House, Cube M4 Business Park Old Gloucester Road, Bristol BS16 1FX, UK","note":"UK-registered company, no media-region statement on the rates page."},"compliance":{"value":"ISO 27001 and ISO 9001 (footer), 100% availability SLA for UK carrier services","source":"https://simwood.com/pricing","read":"2026-10-29","quote":"ISO 27001 · ISO 9001","note":"Rates page also says: 'Unlike other operators, we are not passing on UK origin surcharges'."}}},{"id":"ofcom-rules","company":"Ofcom","product":"CLI Guidance (General Condition C6, calling line identification)","layer":"carrier","homepage":"https://www.ofcom.org.uk/phones-and-broadband/phone-numbers","summary":"The UK regulator's rule that every call must carry a valid, dialable number that identifies the caller, and that bad caller IDs get blocked.","facts":{"cli_who_it_applies_to":{"value":"Any provider of a calling service that uses telephone numbers, including voice over IP, applies from 29 January 2025","source":"https://www.ofcom.org.uk/siteassets/resources/documents/consultations/category-2-6-weeks/276698---further-action-to-tackle--scam-calls/associated-documents/annex-2-cli-guidance-2024-update.pdf","read":"2026-10-29","quote":"Therefore these requirements apply to any provider of a calling service that makes use of telephone numbers from a national or international numbering plan, including voice services using technologies such as IP.","note":"Cover page: 'Published: 29 July 2024 Applies from: 29 January 2025'. Section 1.2 says the requirements are in General Condition C6."},"cli_valid_number":{"value":"Caller ID on every call must be a valid, dialable number that uniquely identifies the caller. An AI phone agent calling out needs a number it is entitled to use as the caller ID.","source":"https://www.ofcom.org.uk/siteassets/resources/documents/consultations/category-2-6-weeks/276698---further-action-to-tackle--scam-calls/associated-documents/annex-2-cli-guidance-2024-update.pdf","read":"2026-10-29","quote":"They also require CPs to ensure that any CLI Data provided with and/or associated with a call includes a valid, dialable telephone number which uniquely identifies the caller.","note":"Section 1.6. The second sentence is a plain-English reading by the researcher, not an Ofcom quote."},"cli_blocking":{"value":"Calls with inaccurate caller ID should be stopped or blocked","source":"https://www.ofcom.org.uk/siteassets/resources/documents/consultations/category-2-6-weeks/276698---further-action-to-tackle--scam-calls/associated-documents/annex-2-cli-guidance-2024-update.pdf","read":"2026-10-29","quote":"Calls with inaccurate CLI Data should be stopped or blocked, to protect recipients from spoofed calls.","note":"Section 1.5. Expect carriers to reject calls that present a caller ID you have no right to use."},"number_allocation":{"value":null,"source":"https://www.ofcom.org.uk/phones-and-broadband/phone-numbers","read":"2026-10-29","quote":"Information about telephone numbers in the UK, including area codes and call costs.","note":"Not opened: Ofcom's General Condition B1 on who may be allocated and adopt numbers. A search result summarised it but We did not read the text, so no claim is made here. The phone-numbers hub page links to Ofcom's numbering data, last updated 28 October 2026."}}}]}