Voice AI BenchmarksBy

DEEPGRAMSPEECH-TO-TEXT114,205 SAMPLES / 30 DAYSLAST RUN SEP 15, 2026, 08:00 UTC

official resource

Nova 3 speech-to-text benchmarks

Nova 3, hosted by Deepgram, measures mean 89 ms time to final segment (6th of 28) and 6.2% word error rate (20th of 30) among STT systems. Results cover the last 30 days. Coval benchmarks hosted endpoints daily using a consistent evaluation methodology. Last measured .

Nova 3 is Deepgram's current general-purpose speech-to-text model.

Time to Final Segment#6 / 28
89ms
Word Error Rate#20 / 30
6.2%
Time to First Token#9 / 26
1418ms

Overview

Nova 3 handles multilingual audio in real time, and the tested endpoint adds diarization, code switching and keyterm controls.

Nova 3 is tested every day on fixed public audio — clean, accented, noisy, reverberant, far-field, clipped and phone-codec speech — for transcription accuracy and streaming latency.

Technical specifications

Made by
Deepgram
Hosted by
Deepgram
Source
Official API
Licensing
Proprietary
Deployment
On-prem
Region
US
Features
Code switching, Diarization, Keyterm biasing, Multilingual, VAD

How Nova 3 ranks

Full STT dashboard
Time to Final Segment — every measured STT modelMilliseconds · lower is better · 30-day averageEvery STT model ranked on Time to Final Segment, with Nova 3 highlighted.
  1. #1Qwen3 ASR 1.7b39 ms
  2. #2Qwen3 ASR Fast46 ms
  3. #3STT RT v557 ms
  4. #4STT 165 ms
  5. #5Parakeet TDT 0.6B v381 ms
  6. #6Nova 389 ms
  7. #7Nova 292 ms
  8. #8Flux Multilingual98 ms
  9. #9Flux99 ms
  10. #10Ink 2122 ms
  11. #11Whisper Large v3via Baseten125 ms
  12. #12Scribe v2 Realtime133 ms
Show all 28 models
  1. #13Universal 3.5 Pro173 ms
  2. #14Grok STT207 ms
  3. #15Defaultvia Speechmatics209 ms
  4. #16Pulse212 ms
  5. #17Defaultvia Gradium246 ms
  6. #18resonant-1264 ms
  7. #19Whisper Large v3via Together AI296 ms
  8. #20Enhanced299 ms
  9. #21Gemini 3.5 Transcribe Live305 ms
  10. #22Voxtral Mini Transcribe Realtime 2602400 ms
  11. #23GPT Realtime Whisper551 ms
  12. #24GPT-4o mini Transcribe690 ms
  13. #25GPT-4o Transcribe754 ms
  14. #26Chirp 3776 ms
  15. #27Solaria 1803 ms
  16. #28Chirp 2873 ms
median of all models · 208 ms
STT models over the last 30 days, ranked on Time to Final Segment.
#ModelHostTTFSWERTTFTSamples
1Qwen3 ASR 1.7bBaseten39 ms931 ms1,256
2Qwen3 ASR FastNari46 ms1748 ms4,411
3STT RT v5Soniox57 ms1529 ms22,468
4STT 1Inworld AI65 ms1400 ms22,796
5Parakeet TDT 0.6B v3Together AI81 ms1215 ms22,448
6Nova 3Deepgram89 ms1418 ms22,815
7Nova 2Deepgram92 ms1419 ms22,693
8Flux MultilingualDeepgram98 ms1163 ms12,211
9FluxDeepgram99 ms1076 ms12,203
10Ink 2Cartesia122 ms1827 ms22,822
11Whisper Large v3Baseten125 ms912 ms1,252
12Scribe v2 RealtimeElevenLabs133 ms2175 ms22,826
13Universal 3.5 ProAssemblyAI173 ms1040 ms20,213
14Grok STTxAI207 ms—22,813
15DefaultSpeechmatics209 ms1428 ms22,836
16PulseSmallest212 ms2011 ms22,816
17DefaultGradium246 ms1980 ms22,799
18resonant-1Reson8264 ms—22,820
19Whisper Large v3Together AI296 ms1338 ms22,705
20EnhancedSpeechmatics299 ms1491 ms22,836
21Gemini 3.5 Transcribe LiveGemini305 ms1663 ms16,363
22Voxtral Mini Transcribe Realtime 2602Mistral400 ms1850 ms22,544
23GPT Realtime WhisperOpenAI551 ms1814 ms22,772
24GPT-4o mini TranscribeOpenAI690 ms—22,790
25GPT-4o TranscribeOpenAI754 ms—22,791
26Chirp 3Google776 ms5974 ms22,835
27Solaria 1Gladia803 ms1802 ms22,399
28Chirp 2Google873 ms6072 ms22,749
—Nemotron 3.5 ASR StreamingTogether AI—1548 ms22,707
—Universal StreamingAssemblyAI—1513 ms20,195
Under-sampled models are excluded; tied models share a place. Dotted WER values split into substitutions, deletions and insertions on hover or tap.

Latency vs accuracy

TTFS vs WEREach point is one measured model · Nova 3 highlighted · 30-day averagesTime to Final Segment against Word Error Rate for every measured STT model, with Nova 3 highlighted.
Qwen3 ASR 1.7b — 39 ms, 4.1%Qwen3 ASR Fast — 46 ms, 3.2%STT RT v5 — 57 ms, 5.7%STT 1 — 65 ms, 4.4%Parakeet TDT 0.6B v3 — 81 ms, 11.2%Nova 2 — 92 ms, 7.9%Flux Multilingual — 98 ms, 8.8%Flux — 99 ms, 6.7%Ink 2 — 122 ms, 4.8%Whisper Large v3 via Baseten — 125 ms, 5.6%Scribe v2 Realtime — 133 ms, 5.5%Universal 3.5 Pro — 173 ms, 3.2%Grok STT — 207 ms, 4.6%Default via Speechmatics — 209 ms, 5.4%Pulse — 212 ms, 5.3%Default via Gradium — 246 ms, 10.0%resonant-1 — 264 ms, 3.4%Whisper Large v3 via Together AI — 296 ms, 8.5%Enhanced — 299 ms, 4.3%Gemini 3.5 Transcribe Live — 305 ms, 3.9%Voxtral Mini Transcribe Realtime 2602 — 400 ms, 6.2%GPT Realtime Whisper — 551 ms, 5.1%GPT-4o mini Transcribe — 690 ms, 4.9%GPT-4o Transcribe — 754 ms, 4.6%Chirp 3 — 776 ms, 4.1%Solaria 1 — 803 ms, 7.1%Chirp 2 — 873 ms, 4.9%Nova 3 — 89 ms, 6.2%

Where the errors come from

WER compositionNova 3's Word Error Rate split by error type · 30-day averageNova 3's WER split into substitutions, deletions and insertions.
  • Nova 36.2%
SubstitutionsDeletionsInsertions

Averages and tail latency

Averages hide slow outliers — these are the distributions behind each figure.

Nova 3 latency distributionMilliseconds · shared axis across metrics · last 30 daysNova 3's latency percentiles per metric: p25–p75 band, p50 tick, whisker to p99.
  • Time to Final Segmentp50 92 ms · p99 173 ms
  • Time to First Tokenp50 997 ms · p99 2985 ms
band p25–p75 · tick p50 · whisker to p99 with p90 and p95 stops
Average and percentile values per metric, with the number of samples behind each row.
MetricAveragep25p50p75p90p95p99Samples
Time to Final Segment89 ms59 ms92 ms104 ms123 ms136 ms173 ms22,815
Word Error Rate6.2%0.0%3.1%8.7%16.6%22.2%46.9%22,853
Time to First Token1418 ms980 ms997 ms1980 ms1989 ms2007 ms2985 ms22,853

Last 30 days

Daily medians from the same measurement runs · gaps are days without qualifying runs.

Time to Final Segment — daily p50Line p50 · band p25–p75 · UTC daysNova 3's daily median Time to Final Segment over the last 30 days.
Nova 3 · Sep 1: 92 msNova 3 · Sep 2: 86 msNova 3 · Sep 3: 92 msNova 3 · Sep 4: 91 msNova 3 · Sep 5: 87 msNova 3 · Sep 6: 86 msNova 3 · Sep 7: 87 msNova 3 · Sep 8: 90 msNova 3 · Sep 9: 91 msNova 3 · Sep 10: 93 msNova 3 · Sep 11: 92 msNova 3 · Sep 12: 84 msNova 3 · Sep 13: 83 msNova 3 · Sep 14: 89 msNova 3 · Sep 15: 82 ms
Word Error Rate — daily averageDaily average · UTC daysNova 3's daily Word Error Rate over the last 30 days.
Nova 3 · Sep 1: 5.1%Nova 3 · Sep 2: 6.5%Nova 3 · Sep 3: 7.0%Nova 3 · Sep 4: 6.4%Nova 3 · Sep 5: 6.3%Nova 3 · Sep 6: 7.1%Nova 3 · Sep 7: 7.5%Nova 3 · Sep 8: 6.4%Nova 3 · Sep 9: 6.7%Nova 3 · Sep 10: 6.8%Nova 3 · Sep 11: 6.8%Nova 3 · Sep 12: 6.7%Nova 3 · Sep 13: 7.1%Nova 3 · Sep 14: 6.4%Nova 3 · Sep 15: 7.7%

Time to Final Segment by dataset

Nova 3 by test conditionMilliseconds · lower is better · best condition firstNova 3's Time to Final Segment on each benchmark dataset.
  1. WildASR accents83 ms
  2. WildASR far-field87 ms
  3. WildASR reverb87 ms
  4. WildASR noise gaps88 ms
  5. PipeCat (production)88 ms
  6. WildASR clean89 ms
  7. WildASR clipping92 ms
  8. WildASR phone codec94 ms
Time to Final Segment per dataset, with the number of samples behind each figure.
DatasetTTFSSamples
LibriSpeech105 ms1
ProductionPipeCat88 ms11,505
AccentsWildASR83 ms1,116
CleanWildASR89 ms4,574
ClippingWildASR92 ms1,139
Far-fieldWildASR87 ms1,137
Noise gapsWildASR88 ms1,142
Phone codecWildASR94 ms1,136
ReverbWildASR87 ms1,085

Strongest condition: WildASR accents at 83 ms · weakest: LibriSpeech at 105 ms.

How fast is Nova 3?

On Deepgram, Nova 3 measures mean 89 ms time to final segment (6th of 28) and mean 1418 ms time to first token (9th of 26). Last measured 2026-09-15.

How accurate is Nova 3?

On Deepgram, Nova 3 measures 6.2% word error rate (20th of 30). Last measured 2026-09-15.

Who hosts Nova 3?

Nova 3 is created by Deepgram and served by Deepgram. Coval measures each hosted endpoint separately.

Limits of this comparison

Coval reports separate results for clean, accented, noisy, reverberant, far-field, clipped and phone-codec audio.

  • Language, endpointing and keyword settings can change production outcomes beyond Coval's reproducible configuration.

Official sources

Results are re-measured daily using fixed datasets and reported over a rolling 30-day window. Same datasets, prompts and metric definitions for every model, measured by Coval’s open-source runner. Full methodology on the overview.

Evaluate your own voice agent

Use Coval to test your production configuration, prompts and calls—not only the public benchmark endpoints.

Book a Demo