Voice AI provider arena
Every part of a phone voice agent, from the platform to the phone line. Performance first, where someone independent has measured it, then what it costs. Every value links to its source and the date we read it.
How they perform, against what they cost
Measured by others- Voice agent platformsNot measured yet
- Speech-to-text13 of 23 measured
- Text-to-speech1 single figure, too few to compare
- Speech-to-speech models5 of 10 measured
- Language models for voice1 single figure, too few to compare
- Phone carriersNo performance measure exists
What you pay against how accurate it is
Measured by othersEach dot has both a list price and a Pipecat figure. Lower and further left is cheaper and better.
Semantic word errors (Pipecat), lower is better
Shaded: cheaper than the median and below the median on errors. Dashed lines: the medians.
Not on this chart. No list price or no Pipecat figure: Deepgram Flux, OpenAI gpt-live-transcribe, OpenAI gpt-transcribe, OpenAI gpt-4o-mini-transcribe, Google Chirp 3, Azure MAI-Transcribe-2, Gladia Solaria, Rev.ai Reverb, Mistral Voxtral Realtime 3, Mistral Voxtral Realtime v26.02, OpenAI Whisper large-v3 / large-v3-turbo, NVIDIA Parakeet TDT 0.6B v3, NVIDIA Canary 1B v2, Kyutai STT 2.6B.
Charted from independent figures: speech-to-text and speech-to-speech models. The other layers show the few independent figures there are, who measures them elsewhere and what each vendor claims, unranked. Our own test round is planned.
Every list price on one scale
List pricesUS dollars a minute, on a log scale, so each darker gridline is ten times the one before. Hover or tap a dot for its source, click it for the full row.
RingTru has not run its own provider tests yet. Nothing here is our measurement. Our first round is planned.
Compare one layer
The best five on an independent measure where one exists. Where too few independent figures exist, the card says so, gives the price range and what we will measure, and ranks nothing.
How the parts fit together
A chained agent passes the call through each part in turn. A speech-to-speech model replaces the middle three with one model that hears and speaks.
Speech-to-speech vs chained agents, the guide- 1Phone carrier The number and the line
- 2Platform Runs the call
- 3Speech-to-text Hears the caller
- 4Language model Decides what to say
- 5Text-to-speech Speaks the reply
What does a minute of call cost?
Pick a part from each layer to add up list prices. The platform fee alone is rarely the whole bill.
Pick at least one part to see a cost.
List prices in US dollars, before tax, free tiers and volume discounts. Speech-to-text is counted for the whole call, text-to-speech at 850 characters a minute of agent speech. For a speech-to-speech model, caller audio is counted for the caller’s share of the call and agent audio for the agent’s. Some vendors also bill silence. Prices sold only in another currency, or with no published rate, are left out.
How to read these numbers
Five rules the whole arena follows.
- Vendor claims sit apart from measurements
- What a company says about its own product is labelled as its claim. Figures measured by someone else, such as a public leaderboard or a paper, sit in their own column with who measured them. The two are never blended into one score.
- Every value has a source and a date
- Each cell links to the page it came from and shows when we read it. Prices change often, so check the date before you rely on one.
- "Not published" is an answer
- Where a company does not publish a figure, the table says so and links to the page we checked, rather than leaving a gap or guessing.
- Clean audio is not a phone call
- Most accuracy figures, independent ones included, come from clean recordings. Phone lines carry 8 kHz audio and accents fare worse on them, so treat every accuracy figure as a best case for a phone agent.
- List prices, not what you will pay
- The arena shows pay-as-you-go prices as printed. Volume deals, free tiers, taxes and minimum spends are left out.
Independent sources
Who else measures these providers, and whether their figures may be shown here.
Figures shown here (11)
- Pipecat STT benchmark (pipecat-ai/stt-benchmark)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.Run by Pipecat
- tau-Voice (arXiv 2603.13686)Task completion (pass@1) of full-duplex voice agents on tau2-bench style tasks, plus latency, responsiveness, interruption and selectivity.Run by Soham Ray
- Mind the Accent Gap: British Accent Robustness in Speech-Driven Financial Voice Assistants (arXiv 2610.06587)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.Run by Aadam Haq
- Responsible ASR: Overcoming Challenges of Foundational Models in Narrow-Band and Low-Resource Settings (arXiv 2606.18659)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.Run by Tejas Godambe
- mu-bench: speech-to-text on 8 kHz phone calls (arXiv 2609.32082)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.Run by Andrea Li
- Full-Duplex-Bench-v3 (arXiv 2604.04847)Tool use by voice agents when people speak with real hesitations: tool choice, argument accuracy, pass@1, turn-taking, latency and interruptions.Run by Lin
- Full-Duplex-Bench v1.5 (arXiv 2507.23159)How speech-to-speech models handle overlapping speech: interruptions, backchannels, side talk and background speech.Run by Lin et al.
- EVA-Bench (arXiv 2605.13841)Task accuracy (EVA-A) and caller experience (EVA-X) of end-to-end voice agents over 213 scenarios, 5 runs each.Run by ServiceNow AI Research. Not a seller of voice models.
- VoiceBench (arXiv 2410.17196)Spoken question answering and instruction following for voice assistants, overall score out of 100.Run by Chen et al. Academic
- T5Gemma-TTS technical report (arXiv 2604.01760)Naturalness (UTMOS22, a predicted listener score out of 5), word error rate and speaker similarity on LibriSpeech English.Run by Arata and Kurihara
- Berkeley Function Calling Leaderboard (BFCL v4)How often a model calls the right tool with the right arguments, run by BFCL on its own hosting.Run by UC Berkeley Gorilla project. Academic.
Linked only, their terms do not allow copying (7)
- Hugging Face Open ASR LeaderboardAverage 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).Run by Hugging Face with NVIDIA
- Coval Voice AI Benchmarks (benchmarks.coval.ai)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).Run by Coval
- Artificial Analysis Speech to TextSTT accuracy (AA-WER index), speed and price per provider, methodology at /methodology/speech-to-text.Run by Artificial Analysis
- Artificial Analysis Text to SpeechTTS quality via its Speech Arena (blind votes), latency and price.Run by Artificial Analysis
- Artificial Analysis Speech to SpeechSpeech-to-speech models: credits tau-Voice, Full Duplex Bench and Big Bench Audio, methodology at /methodology/speech-to-speech-benchmarking.Run by Artificial Analysis
- TTS Arena V2 (blind listening votes)ELO rating from blind head-to-head votes on which voice sounds more human.Run by TTS-AGI community
- Picovoice speech-to-text-benchmarkEnglish 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.Vendor-run, read as a claim: Picovoice
Download the data
Every row, with its sources, quotes and read dates.
Your agent, tested on real calls.
- test calls
- 506
- passed
- 96.8%
- median reply
- 0.94 s
Holds the new day when a caller changes their mind.