Speech-to-text compared
Speech-to-text is how a voice agent hears the caller. In independent tests the best gets under 1% of words wrong in a way that changes the meaning. On 8 kHz phone calls the best gets 2.8% of caller turns wrong. List prices run from about $0.002 to about $0.017 a minute.
Word errors that change the meaning, measured by Pipecat. Lower is better.
- Best
- 0.89%AssemblyAI Universal-3.6 Pro Realtime
- Median
- 1.87%Of 10 measured by Pipecat
- Weakest
- 4.49%Mistral Voxtral Realtime v26.02
- Measured by others
- 13 of 23Have an independent figure
23 compared, read 29 October 2026 or later. Click a bar to open its row, a column heading to sort, ↗ for the source. Tick Compare on up to four rows.
Showing 23 of 23
Word errors that change the meaning, measured by Pipecat
Measured by othersSemantic word error rate over 1,000 recorded voice-agent turns, from the open Pipecat speech-to-text benchmark. Lower is better.
Not on this chart. Not measured by Pipecat: 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, OpenAI Whisper large-v3 / large-v3-turbo, NVIDIA Parakeet TDT 0.6B v3, NVIDIA Canary 1B v2, Kyutai STT 2.6B.
- Accuracy, bestAssemblyAI Universal-3.6 Pro Realtime 0.89%
- Phone calls, bestAzure Speech real-time 2.8%
- Speed, bestDeepgram Nova-3 247 ms
- Price, lowestSoniox stt-rt-v5 $0.002
Bar length is the figure itself, so the shortest bar is best in every column.
| Pipecat, lower is better | mu-bench, lower is better | Pipecat, lower is better | US$ a minute | Checks✓ yes, ✕ no, ~ partly | Measured by otherssingle figures | |
|---|---|---|---|---|---|---|
| AssemblyAI Universal-3.6 Pro Realtime | 0.89%Best | Not measured by mu-bench | 307 ms | $0.0075 | Not published | None found |
| Microsoft Azure Azure Speech real-time speech to text | 0.98% | 2.8%Best | 1,016 ms | $0.0167 | British English | None found |
| Speechmatics Linden 1 (Agent STT) | 1.11% | Not measured by mu-bench | 369 ms | $0.005 | British English | None found |
| Soniox stt-rt-v5 | 1.11% | Not measured by mu-bench | 260 ms | $0.002Lowest | Not published | None found |
| Deepgram Nova-3 | 1.32% | 3.9% | 247 msBest | $0.0077 | British English | None found |
| OpenAI gpt-realtime-whisper | 2.41% | Not measured by mu-bench | 740 ms | $0.017 | Not published | None found |
| AssemblyAI Universal-Streaming (English / Multilingual) | 2.81% | Not measured by mu-bench | 256 ms | $0.0025 | Not published | None found |
| ElevenLabs Scribe v2 Realtime | 3.05% | Not measured by mu-bench | 281 ms | $0.0065 | Not published | None found |
| OpenAI gpt-4o-transcribe | 3.1% | Not measured by mu-bench | 637 ms | $0.006 | Not published | None found |
| Mistral AI Voxtral Mini Transcribe Realtime (v26.02, open weights)Open weights | 4.49% | Not measured by mu-bench | 525 ms | Not published | Not published | None found |
| Deepgram Flux | Not measured by Pipecat | Not measured by mu-bench | Not measured by Pipecat | $0.0077 | British English | None found |
| OpenAI gpt-live-transcribe | Not measured by Pipecat | Not measured by mu-bench | Not measured by Pipecat | $0.017 | Not published | None found |
| OpenAI gpt-transcribe | Not measured by Pipecat | Not measured by mu-bench | Not measured by Pipecat | $0.0045 | Not published | None found |
| OpenAI gpt-4o-mini-transcribe | Not measured by Pipecat | 4% | Not measured by Pipecat | $0.003 | Not published | None found |
| Google Cloud Speech-to-Text V2 (Chirp 3) | Not measured by Pipecat | 3.2% | Not measured by Pipecat | $0.016 | British English | None found |
| Microsoft Azure MAI-Transcribe-2-Streaming (preview) | Not measured by Pipecat | Not measured by mu-bench | Not measured by Pipecat | Not published | Not published | None found |
| Gladia Solaria (real-time) | Not measured by Pipecat | Not measured by mu-bench | Not measured by Pipecat | Not comparable | Not published | None found |
| Rev.ai Reverb (streaming) | Not measured by Pipecat | Not measured by mu-bench | Not measured by Pipecat | Not published | Not published | None found |
| Mistral AI Voxtral Mini Transcribe Realtime 3 | Not measured by Pipecat | Not measured by mu-bench | Not measured by Pipecat | $0.008 | Not published | None found |
| OpenAI (open source) Whisper large-v3 / large-v3-turboOpen weights | Not measured by Pipecat | Not measured by mu-bench | Not measured by Pipecat | Not published | Not published | |
| NVIDIA Parakeet TDT 0.6B v3Open weights | Not measured by Pipecat | Not measured by mu-bench | Not measured by Pipecat | Not published | Not published | None found |
| NVIDIA Canary 1B v2Open weights | Not measured by Pipecat | Not measured by mu-bench | Not measured by Pipecat | Not published | Not published | None found |
| Kyutai Kyutai STT 2.6B (English)Open weights | Not measured by Pipecat | Not measured by mu-bench | Not measured by Pipecat | Not published | Not published | None found |
Vendor claims, data regions and compliance are in each row’s Details, in the vendor’s own words. Claims are never charted or ranked. RingTru has measured none of these yet. Add up a whole call in the calculator.
How to read these numbers
- Word error rate
- The share of words the model got wrong, counting substitutions, missed words and added words. Lower is better. A figure only means something next to the audio it was measured on: clean read speech scores far better than a phone call.
- Vendor claims and independent figures
- A vendor that benchmarks itself chooses the test set. Independent figures, such as the Hugging Face Open ASR Leaderboard, run every model on the same public sets, though still mostly on clean audio. We show both, labelled, and never blend them.
- Streaming latency
- For a phone agent, what matters is the time from the caller finishing a word to the final transcript arriving. Vendors measure this in different ways, and few say which.
- Phone audio
- Phone lines carry 8 kHz audio, half the band most models are trained on. Accuracy drops, and for regional accents it can drop sharply. The arena does not yet have independent phone-audio figures for most models.
Go deeper
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.