Voice AI BenchmarksBy

GRADIUMSPEECH-TO-TEXT136,820 SAMPLES / 30 DAYSLAST RUN NOV 2, 2026, 09:00 UTC

Default speech-to-text benchmarks

Default, hosted by Gradium, measures mean 261 ms time to final segment (21st of 28) and 8.4% word error rate (28th of 30) among STT systems. Results cover the last 30 days. Coval benchmarks hosted endpoints daily using a consistent evaluation methodology. Last measured .

Coval has active speech-to-text measurements for Default, created by Gradium and served through Gradium's own API.

Time to Final Segment#21 / 28
261ms
Word Error Rate#28 / 30
8.4%
Time to First Token#22 / 28
2005ms

Overview

Default 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
Gradium
Hosted by
Gradium
Source
Official API
Licensing
Proprietary
Deployment
Cloud
Region
US
Features
Multilingual, VAD

How Default 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 Default highlighted.
  1. #1Qwen3 ASR 1.7b35 ms
  2. #2STT RT v544 ms
  3. #3Qwen3 ASR Fast50 ms
  4. #4STT 170 ms
  5. #5Nova 385 ms
  6. #6Nova 287 ms
  7. #7Parakeet TDT 0.6B v388 ms
  8. #8Flux Multilingual100 ms
  9. #9Flux101 ms
  10. #10Universal 3.6 Pro114 ms
  11. #11Scribe v2 Realtime118 ms
  12. #12Whisper Large v3via Baseten126 ms
  13. #13Universal 3.5 Pro126 ms
  14. #14Ink 2146 ms
  15. #15Realtime Global159 ms
  16. #16Whisper Large v3via Together AI174 ms
  17. #17Grok STT180 ms
  18. #18Pulse203 ms
  19. #19resonant-1243 ms
  20. #20Linden 1254 ms
  21. #21Default261 ms
Show all 28 models
  1. #22Gemini 3.5 Transcribe Live383 ms
  2. #23Daytona STT384 ms
  3. #24Solaria 1414 ms
  4. #25Voxtral Mini Transcribe Realtime 2602436 ms
  5. #26GPT Realtime Whisper563 ms
  6. #27Chirp 3807 ms
  7. #28Chirp 2895 ms
median of all models · 152 ms
STT models over the last 30 days, ranked on Time to Final Segment.
#ModelHostTTFSWERTTFTSamples
1Qwen3 ASR 1.7bBaseten35 ms936 ms13,678
2STT RT v5Soniox44 ms1525 ms27,365
3Qwen3 ASR FastNari Labs50 ms1755 ms27,371
4STT 1Inworld AI70 ms1026 ms27,334
5Nova 3Deepgram85 ms1444 ms27,364
6Nova 2Deepgram87 ms1427 ms27,311
7Parakeet TDT 0.6B v3Together AI88 ms1197 ms15,445
8Flux MultilingualDeepgram100 ms1174 ms27,365
9FluxDeepgram101 ms1097 ms27,371
10Universal 3.6 ProAssemblyAI114 ms1079 ms27,335
11Scribe v2 RealtimeElevenLabs118 ms2151 ms27,366
12Whisper Large v3Baseten126 ms914 ms13,674
13Universal 3.5 ProAssemblyAI126 ms1019 ms27,373
14Ink 2Cartesia146 ms2034 ms27,369
15Realtime GlobalReson8159 ms2060 ms22,469
16Whisper Large v3Together AI174 ms1210 ms15,513
17Grok STTSpaceXAI180 ms—27,332
18PulseSmallest203 ms1988 ms27,366
19resonant-1Reson8243 ms—27,372
20Linden 1Speechmatics254 ms1528 ms27,368
21DefaultGradium261 ms2005 ms27,364
22Gemini 3.5 Transcribe LiveGemini383 ms1542 ms27,270
23Daytona STTGuava384 ms5651 ms13,133
24Solaria 1Gladia414 ms1608 ms27,306
25Voxtral Mini Transcribe Realtime 2602Mistral436 ms1867 ms20,731
26GPT Realtime WhisperOpenAI563 ms1822 ms26,862
27Chirp 3Google807 ms6007 ms27,330
28Chirp 2Google895 ms6097 ms27,368
—Nemotron 3.5 ASR StreamingTogether AI—1557 ms15,422
—Universal StreamingAssemblyAI—1516 ms27,339
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 · Default highlighted · 30-day averagesTime to Final Segment against Word Error Rate for every measured STT model, with Default highlighted.
Qwen3 ASR 1.7b — 35 ms, 2.9%STT RT v5 — 44 ms, 4.8%Qwen3 ASR Fast — 50 ms, 2.6%STT 1 — 70 ms, 4.4%Nova 3 — 85 ms, 5.2%Nova 2 — 87 ms, 6.6%Parakeet TDT 0.6B v3 — 88 ms, 10.9%Flux Multilingual — 100 ms, 7.4%Flux — 101 ms, 5.7%Universal 3.6 Pro — 114 ms, 2.2%Scribe v2 Realtime — 118 ms, 4.2%Whisper Large v3 via Baseten — 126 ms, 4.5%Universal 3.5 Pro — 126 ms, 2.4%Ink 2 — 146 ms, 3.4%Realtime Global — 159 ms, 2.6%Whisper Large v3 via Together AI — 174 ms, 7.2%Grok STT — 180 ms, 4.1%Pulse — 203 ms, 4.6%resonant-1 — 243 ms, 2.6%Linden 1 — 254 ms, 3.8%Gemini 3.5 Transcribe Live — 383 ms, 2.9%Daytona STT — 384 ms, 4.8%Solaria 1 — 414 ms, 6.2%Voxtral Mini Transcribe Realtime 2602 — 436 ms, 5.3%GPT Realtime Whisper — 563 ms, 4.0%Chirp 3 — 807 ms, 3.2%Chirp 2 — 895 ms, 3.9%Default — 261 ms, 8.4%

Where the errors come from

WER compositionDefault's Word Error Rate split by error type · 30-day averageDefault's WER split into substitutions, deletions and insertions.
  • Default8.4%
SubstitutionsDeletionsInsertions

Averages and tail latency

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

Default latency distributionMilliseconds · shared axis across metrics · last 30 daysDefault's latency percentiles per metric: p25–p75 band, p50 tick, whisker to max.
  • Time to Final Segmentp50 246 ms · Max 22223 ms
  • Time to First Tokenp50 1964 ms · Max 19283 ms
band p25–p75 · tick p50 · whisker to max with p90 and p95 stops
Average and percentile values per metric, with the number of samples behind each row.
MetricAveragep25p50p75p90p95MaxSamples
Time to Final Segment261 ms221 ms246 ms288 ms314 ms336 ms22223 ms27,364
Word Error Rate8.4%0.0%5.0%12.9%24.0%33.3%100.0%27,364
Time to First Token2005 ms1655 ms1964 ms2247 ms2639 ms2842 ms19283 ms27,364

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 daysDefault's daily median Time to Final Segment over the last 30 days.
Default · Oct 3: 241 msDefault · Oct 4: 233 msDefault · Oct 5: 253 msDefault · Oct 6: 253 msDefault · Oct 7: 266 msDefault · Oct 8: 243 msDefault · Oct 9: 249 msDefault · Oct 10: 256 msDefault · Oct 11: 256 msDefault · Oct 12: 255 msDefault · Oct 13: 256 msDefault · Oct 14: 257 msDefault · Oct 15: 249 msDefault · Oct 16: 250 msDefault · Oct 17: 251 msDefault · Oct 18: 252 msDefault · Oct 19: 251 msDefault · Oct 20: 245 msDefault · Oct 21: 263 msDefault · Oct 22: 261 msDefault · Oct 23: 243 msDefault · Oct 24: 281 msDefault · Oct 25: 263 msDefault · Oct 26: 263 msDefault · Oct 27: 285 msDefault · Oct 28: 244 msDefault · Oct 29: 244 msDefault · Oct 30: 264 msDefault · Oct 31: 285 msDefault · Nov 1: 262 msDefault · Nov 2: 245 ms
Word Error Rate — daily averageDaily average · UTC daysDefault's daily Word Error Rate over the last 30 days.
Default · Oct 3: 10.1%Default · Oct 4: 9.6%Default · Oct 5: 10.7%Default · Oct 6: 8.7%Default · Oct 7: 9.9%Default · Oct 8: 9.3%Default · Oct 9: 8.4%Default · Oct 10: 8.5%Default · Oct 11: 10.2%Default · Oct 12: 8.7%Default · Oct 13: 9.8%Default · Oct 14: 9.8%Default · Oct 15: 10.4%Default · Oct 16: 9.6%Default · Oct 17: 9.6%Default · Oct 18: 8.8%Default · Oct 19: 9.7%Default · Oct 20: 9.5%Default · Oct 21: 8.8%Default · Oct 22: 9.4%Default · Oct 23: 8.9%Default · Oct 24: 11.4%Default · Oct 25: 8.1%Default · Oct 26: 9.6%Default · Oct 27: 9.4%Default · Oct 28: 9.3%Default · Oct 29: 10.7%Default · Oct 30: 9.3%Default · Oct 31: 9.3%Default · Nov 1: 9.5%Default · Nov 2: 8.0%

Time to Final Segment by dataset

Default by test conditionMilliseconds · lower is better · best condition firstDefault's Time to Final Segment on each benchmark dataset.
  1. PipeCat (production)240 ms
  2. WildASR clean263 ms
  3. WildASR clipping276 ms
  4. WildASR noise gaps281 ms
  5. WildASR far-field294 ms
  6. WildASR phone codec295 ms
  7. WildASR accents361 ms
Time to Final Segment per dataset, with the number of samples behind each figure.
DatasetTTFSSamples
ProductionPipeCat240 ms14,408
AccentsWildASR361 ms1,441
CleanWildASR263 ms5,760
ClippingWildASR276 ms1,442
Far-fieldWildASR294 ms1,435
Noise gapsWildASR281 ms1,439
Phone codecWildASR295 ms1,439

Strongest condition: PipeCat (production) at 240 ms · weakest: WildASR accents at 361 ms.

How fast is Default?

On Gradium, Default measures mean 261 ms time to final segment (21st of 28) and mean 2005 ms time to first token (22nd of 28). Last measured 2026-11-02.

How accurate is Default?

On Gradium, Default measures 8.4% word error rate (28th of 30). Last measured 2026-11-02.

Who hosts Default?

Default is created by Gradium and served by Gradium. Coval measures each hosted endpoint separately.

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