Voice AI BenchmarksBy

DATASETSTTWILDASR / FLEURSACTIVE

manifest

WildASR far-field speech-to-text benchmark dataset

284 clips re-rendered as if recorded at a distance from the microphone.

Items
284
fixed public inputs
Models measured
31
last 30 days
Current leader · WER#1 / 31
3.1%
Universal 3.6 Provia AssemblyAI

How models rank on WildASR far-field

Full STT dashboard
Word Error Rate on WildASR far-fieldPercent · lower is better · 30-day average on this datasetEvery STT model measured on the WildASR far-field dataset, ranked on Word Error Rate.
  1. #1Universal 3.6 Pro3.1%
  2. #2Universal 3.5 Pro3.3%
  3. #3Qwen3 ASR Fast4.2%
  4. #4Qwen3 ASR 1.7b4.3%
  5. #5GPT-4o Transcribe4.4%
  6. #6resonant-14.5%
  7. #7GPT-4o mini Transcribe5.7%
  8. #8Whisper Large v3via Baseten5.9%
  9. #9Linden 16.0%
  10. #10Chirp 36.0%
  11. #11Scribe v2 Realtime6.3%
  12. #12Gemini 3.5 Transcribe Live6.4%
Show all 31 models
  1. #13Chirp 26.5%
  2. #14Ink 26.9%
  3. #15Grok STT7.4%
  4. #16Pulse8.3%
  5. #17STT RT v58.5%
  6. #18STT 18.9%
  7. #19GPT Realtime Whisper9.7%
  8. #20Voxtral Mini Transcribe Realtime 26029.7%
  9. #21Nova 310.0%
  10. #22Solaria 110.4%
  11. #23Universal Streaming10.8%
  12. #24Whisper Large v3via Together AI11.3%
  13. #25Flux11.6%
  14. #26Daytona STT11.8%
  15. #27Parakeet TDT 0.6B v312.1%
  16. #28Nova 214.1%
  17. #29Flux Multilingual14.8%
  18. #30Default20.6%
  19. #31Nemotron 3.5 ASR Streaming22.2%
median of all models · 8.3%
STT models on the WildASR far-field dataset over the last 30 days, ranked on Word Error Rate.
#ModelHostTTFSWERTTFTSamples
1Universal 3.6 ProAssemblyAI138 ms1085 ms1,275
2Universal 3.5 ProAssemblyAI188 ms949 ms1,406
3Qwen3 ASR FastNari Labs46 ms1751 ms1,406
4Qwen3 ASR 1.7bBaseten84 ms887 ms720
5GPT-4o TranscribeOpenAI790 ms—1,404
6resonant-1Reson8252 ms—1,406
7GPT-4o mini TranscribeOpenAI745 ms—1,405
8Whisper Large v3Baseten137 ms861 ms720
9Linden 1Speechmatics304 ms1603 ms1,400
10Chirp 3Google774 ms6297 ms1,405
11Scribe v2 RealtimeElevenLabs124 ms2130 ms1,405
12Gemini 3.5 Transcribe LiveGemini316 ms1719 ms1,403
13Chirp 2Google917 ms6490 ms1,406
14Ink 2Cartesia151 ms1965 ms1,406
15Grok STTSpaceXAI226 ms—1,402
16PulseSmallest214 ms2023 ms1,406
17STT RT v5Soniox43 ms1507 ms1,242
18STT 1Inworld AI71 ms1184 ms1,303
19GPT Realtime WhisperOpenAI599 ms1852 ms1,404
20Voxtral Mini Transcribe Realtime 2602Mistral406 ms1833 ms1,381
21Nova 3Deepgram88 ms1269 ms1,372
22Solaria 1Gladia639 ms1722 ms1,377
23Universal StreamingAssemblyAI—1558 ms1,399
24Whisper Large v3Together AI194 ms1192 ms1,397
25FluxDeepgram100 ms811 ms1,372
26Daytona STTGuava404 ms6508 ms288
27Parakeet TDT 0.6B v3Together AI95 ms1144 ms1,395
28Nova 2Deepgram89 ms1315 ms1,369
29Flux MultilingualDeepgram103 ms1089 ms1,372
30DefaultGradium279 ms1942 ms1,402
31Nemotron 3.5 ASR StreamingTogether AI—1594 ms1,401
Under-sampled models are excluded; tied models share a place. Dotted WER values split into substitutions, deletions and insertions on hover or tap.

Inside the dataset

284 fixed public inputs — every model is tested on exactly these.

  • CLIP 1 · 14.9s

    “"he [wales] basically lied to us from the start. first by acting as if this was for legal reasons. second by pretending he was listening to us right up to his art deletion.”

  • CLIP 2 · 12.6s

    “a car bomb detonated at police headquarters in gaziantep turkey yesterday morning killed two police officers and injured more than twenty other people”

  • CLIP 3 · 11.5s

    “a civilization is a singular culture shared by a significant large group of people who live and work co-operatively a society”

3 of 284 clips, streamed from the public benchmark storage bucket.

Source: bosonai/WildASR environment_degradation__en__fleurs_far_field_en · License: Apache-2.0 (WildASR; audio derived from FLEURS, CC-BY-4.0)

WER change from the clean baseline

Same provider and model on this condition vs the WildASR clean dataset; positive = more recognition error.

WildASR far-field vs WildASR cleanPercentage points of WER · diverging from no changePer-model WER change between WildASR clean and WildASR far-field.
zero = no change from the clean baseline · right = more error under this condition
Per-model WER change between WildASR clean and WildASR far-field.
ModelHostClean WERCondition WERChange
Universal 3.6 ProAssemblyAI2.5%3.1%+0.6%
Universal 3.5 ProAssemblyAI2.8%3.3%+0.4%
Qwen3 ASR FastNari Labs3.5%4.2%+0.7%
Qwen3 ASR 1.7bBaseten3.7%4.3%+0.6%
GPT-4o TranscribeOpenAI2.8%4.4%+1.6%
resonant-1Reson83.2%4.5%+1.3%
GPT-4o mini TranscribeOpenAI3.2%5.7%+2.5%
Whisper Large v3Baseten4.3%5.9%+1.7%
Linden 1Speechmatics4.7%6.0%+1.3%
Chirp 3Google3.4%6.0%+2.7%
Scribe v2 RealtimeElevenLabs4.0%6.3%+2.3%
Gemini 3.5 Transcribe LiveGemini3.1%6.4%+3.3%
Chirp 2Google4.6%6.5%+1.9%
Ink 2Cartesia4.5%6.9%+2.4%
Grok STTSpaceXAI4.1%7.4%+3.3%
PulseSmallest5.1%8.3%+3.2%
STT RT v5Soniox6.1%8.5%+2.3%
STT 1Inworld AI8.1%8.9%+0.9%
GPT Realtime WhisperOpenAI4.1%9.7%+5.6%
Voxtral Mini Transcribe Realtime 2602Mistral4.6%9.7%+5.1%
Nova 3Deepgram6.1%10.0%+3.9%
Solaria 1Gladia6.0%10.4%+4.4%
Universal StreamingAssemblyAI6.1%10.8%+4.6%
Whisper Large v3Together AI6.7%11.3%+4.6%
FluxDeepgram5.4%11.6%+6.2%
Daytona STTGuava5.2%11.8%+6.6%
Parakeet TDT 0.6B v3Together AI10.7%12.1%+1.4%
Nova 2Deepgram6.9%14.1%+7.3%
Flux MultilingualDeepgram7.2%14.8%+7.6%
DefaultGradium7.8%20.6%+12.8%
Nemotron 3.5 ASR StreamingTogether AI10.5%22.2%+11.7%

What it tests

This condition simulates greater distance between the speaker and microphone. It measures the effect of lower signal level and room acoustics separately from the dedicated reverb condition.

This condition is derived from the WildASR environment family and FLEURS read speech. Each paired condition uses the same utterance and transcript as the WildASR clean baseline, so the WER change isolates the effect of the test condition.

Metrics reported for this dataset

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