Speech-to-Text (STT) benchmark — by Coval
An independent benchmark of speech-to-text providers on Time to Final Segment (TTFS) and Word Error Rate (WER), measured under production-realistic conditions by Coval. Fastest is parakeet-tdt-0.6b-v3 (46 ms median TTFS), lowest word error rate is inworld-stt-1 (2.6%). Mirrored here with attribution; last 7d.
Run by Coval, a voice-AI evaluation platform
This benchmark is produced by Coval, a voice-AI evaluation platform used by teams like Perplexity, Zoom, ServiceNow, and GEICO to test production voice agents. Coval builds testing and evaluation infrastructure for voice AI — it does not sell STTmodels of its own — so it measures every provider as a neutral third party rather than benchmarking its own product. Every number here is Coval's, measured on a pinned dataset under production-realistic conditions; the methodology, runner code, and inputs are open-source (Apache-2.0) and reproducible by re-running the suite. Openbenchmarks mirrors these results daily with attribution.
Speech-to-Text models ranked by Time to Final Segment
Every model runs the same pinned dataset under the same conditions. TTFS (median) is the headline latency; WER is accuracy (lower is better on both). Ranked fastest-first, last 7d:
| # | Model | TTFS median | TTFS p95 | WER | Samples |
|---|---|---|---|---|---|
| 1 | parakeet-tdt-0.6b-v3 Together | 46 ms | 77 ms | 7.8% | 600 |
| 2 | stt-rt-v5 Soniox | 60 ms | 83 ms | 3.9% | 3,360 |
| 3 | stt-rt-v4 Soniox | 60 ms | 86 ms | 3.9% | 3,360 |
| 4 | nova-3 Deepgram | 98 ms | 154 ms | 4.3% | 3,300 |
| 5 | nova-2 Deepgram | 98 ms | 175 ms | 5.1% | 3,303 |
| 6 | whisper-large-v3 Together | 108 ms | 234 ms | 4.7% | 600 |
| 7 | ink-2 Cartesia | 112 ms | 155 ms | 3.8% | 3,351 |
| 8 | scribe_v2_realtime Elevenlabs | 145 ms | 253 ms | 3.7% | 3,359 |
| 9 | default Azure | 166 ms | 236 ms | 4.5% | 1,010 |
| 10 | inworld-stt-1 Inworld | 167 ms | 243 ms | 2.6% | 1,958 |
| 11 | default Speechmatics | 218 ms | 267 ms | 4.6% | 3,360 |
| 12 | default Gradium | 271 ms | 322 ms | 4.3% | 3,360 |
| 13 | grok-stt Xai | 271 ms | 354 ms | 2.8% | 3,330 |
| 14 | pulse Smallest | 285 ms | 315 ms | 3.7% | 3,360 |
| 15 | voxtral-mini-transcribe-realtime-2602 Mistral | 299 ms | 457 ms | 3.4% | 1,989 |
| 16 | universal-3-pro Assemblyai | 333 ms | 816 ms | 3.2% | 1,370 |
| 17 | solaria-1 Gladia | 365 ms | 2,229 ms | 4.0% | 2,817 |
| 18 | universal-3.5-pro Assemblyai | 386 ms | 666 ms | 2.7% | 3,360 |
| 19 | enhanced Speechmatics | 516 ms | 792 ms | 4.1% | 3,360 |
| 20 | gpt-4o-mini-transcribe Openai | 631 ms | 1,237 ms | 2.7% | 3,346 |
| 21 | gpt-realtime-whisper Openai | 667 ms | 823 ms | 3.3% | 3,347 |
| 22 | chirp_2 | 728 ms | 954 ms | 3.8% | 1,467 |
| 23 | gpt-4o-transcribe Openai | 767 ms | 1,179 ms | 2.9% | 3,352 |
| 24 | chirp_3 | 780 ms | 979 ms | 3.4% | 1,469 |
| 25 | universal-streaming Assemblyai | — | — | 5.5% | 3,360 |
| 26 | universal-streaming-multilingual Assemblyai | — | — | 8.6% | 3,360 |
| 27 | flux-general-en Deepgram | — | — | 5.3% | 3,356 |
| 28 | flux-general-multi Deepgram | — | — | 6.2% | 3,360 |
| 29 | reverb Revai | — | — | 6.9% | 1,010 |
| 30 | nemotron-3.5-asr-streaming-0.6b Together | — | — | 12.4% | 584 |
| 31 | nemotron-3-asr-streaming-0.6b Together | — | — | 9.8% | 470 |
Numbers are point-in-time against Coval's pinned dataset and refresh continuously — they don't generalize indefinitely. Full charts, distributions, and windows (24h / 7d / 30d) at benchmarks.coval.ai →