Skip to content
ZenVoice
Download

Benchmarks

Real Speech
Real Numbers

12 speech models on 1.4K clips of real dictation, conversation and reading.

Accuracy and Speed

Each logo is one model, placed by its WER and speed.

Test Set
ZenVoice Speech Bench V1 ↗
Audio
1,382 clips · 3.2 hours
Mac
MacBook Pro M5 · 24 GB
Run
Oct 27, 2026
6%
7%
8%
9%
10%
11%
12%
13%
14%
15%
16%
17%
3×
5×
7×
10×
15×
20×
30×
50×
70×
100×
150×
Better ↗
Parakeet Ultra
Parakeet TDT V3
Parakeet Unified
Parakeet TDT V2
Cohere Transcribe
Qwen3-ASR 0.6B
Voxtral Realtime
Whisper Large V3 Turbo
Apple Speech Analyzer
Whisper Small
Nemotron 3.5
Omnilingual ASR
Speed →Lower WER →
The exact numbers are in the cards below.

Scroll sideways for the whole chart →

Accuracy vs Speed

  • Parakeet UltraRecommendedWER7.4%Speed71×
  • Parakeet TDT V3WER7.7%Speed70×
  • Parakeet UnifiedWER8.1%Speed108×
  • Parakeet TDT V2WER8.1%Speed75×
  • Cohere TranscribeWER8.3%Speed11×
  • Qwen3-ASR 0.6BWER8.4%Speed35×
  • Voxtral RealtimeWER9.5%Speed4×
  • Whisper Large V3 TurboWER9.7%Speed10×
  • Apple Speech AnalyzerWER10.0%Speed60×
  • Whisper SmallWER11.1%Speed32×
  • Nemotron 3.5WER12.8%Speed47×
  • Omnilingual ASRWER15.6%Speed9×

Speech Engines

  • Parakeet UltraRecommended7.4%
  • Parakeet TDT V37.7%
  • Parakeet Unified8.1%
  • Parakeet TDT V28.1%
  • Cohere Transcribe8.3%
  • Qwen3-ASR 0.6B8.4%
  • Voxtral Realtime9.5%
  • Whisper Large V3 Turbo9.7%
  • Apple Speech Analyzer10.0%
  • Whisper Small11.1%
  • Nemotron 3.512.8%
  • Omnilingual ASR15.6%

WER counts words a model got wrong, with missed clips as every word missed; formatted WER also counts capitals and punctuation. Every clip, transcript and score is on Hugging Face ↗.

Formatting

WER after formatting: words a model dropped, changed or invented.

Test Set
Kanso Dictation Bench V1
Dictations
1,200
Mac
Apple M5 · 24 GB
Run
Oct 27, 2026 · Sumi, Fude Nov 7, 2026

  • Kanso Sumi

    0.28% WER · 74.5% exact · 6 ms

  • Kanso Fude

    0.30% WER · 74.6% exact · 9 ms

  • Kanso Satori 1.0Legacy

    0.43% WER · 69.6% exact · 0.47 s

  • Qwen3-1.7B, stock

    29.12% WER · 22.1% exact · 0.28 s

  • Kanso Satori 0.9

    34.78% WER · 39.9% exact · 0.48 s

  • Qwen3-4B Instruct, stock

    53.54% WER · 46.3% exact · 0.52 s

Unformatted, the transcript as it came in scores 21.14%. Exact match is the share of outputs identical to the reference. Sumi and Fude run through ZenVoice’s formatting pass; the others are the model alone. Categories, examples and the dataset are on the Kanso page →.

See every model →Meet Kanso