benchmarks/speech-to-text by coval
voice AI benchmark · by Coval

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.

synced from Coval2026-07-12 01:18 UTC · run 4109 · dataset 82cfb82d84c6 · methodology → · benchmarks.coval.ai → · open-source runner — re-run it yourself →

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:

#ModelTTFS medianTTFS p95WERSamples
1parakeet-tdt-0.6b-v3
Together
46 ms77 ms7.8%600
2stt-rt-v5
Soniox
60 ms83 ms3.9%3,360
3stt-rt-v4
Soniox
60 ms86 ms3.9%3,360
4nova-3
Deepgram
98 ms154 ms4.3%3,300
5nova-2
Deepgram
98 ms175 ms5.1%3,303
6whisper-large-v3
Together
108 ms234 ms4.7%600
7ink-2
Cartesia
112 ms155 ms3.8%3,351
8scribe_v2_realtime
Elevenlabs
145 ms253 ms3.7%3,359
9default
Azure
166 ms236 ms4.5%1,010
10inworld-stt-1
Inworld
167 ms243 ms2.6%1,958
11default
Speechmatics
218 ms267 ms4.6%3,360
12default
Gradium
271 ms322 ms4.3%3,360
13grok-stt
Xai
271 ms354 ms2.8%3,330
14pulse
Smallest
285 ms315 ms3.7%3,360
15voxtral-mini-transcribe-realtime-2602
Mistral
299 ms457 ms3.4%1,989
16universal-3-pro
Assemblyai
333 ms816 ms3.2%1,370
17solaria-1
Gladia
365 ms2,229 ms4.0%2,817
18universal-3.5-pro
Assemblyai
386 ms666 ms2.7%3,360
19enhanced
Speechmatics
516 ms792 ms4.1%3,360
20gpt-4o-mini-transcribe
Openai
631 ms1,237 ms2.7%3,346
21gpt-realtime-whisper
Openai
667 ms823 ms3.3%3,347
22chirp_2
Google
728 ms954 ms3.8%1,467
23gpt-4o-transcribe
Openai
767 ms1,179 ms2.9%3,352
24chirp_3
Google
780 ms979 ms3.4%1,469
25universal-streaming
Assemblyai
5.5%3,360
26universal-streaming-multilingual
Assemblyai
8.6%3,360
27flux-general-en
Deepgram
5.3%3,356
28flux-general-multi
Deepgram
6.2%3,360
29reverb
Revai
6.9%1,010
30nemotron-3.5-asr-streaming-0.6b
Together
12.4%584
31nemotron-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 →