ZenVoiceModels
17 Models
Zero Servers
A voice model turns your speech into text. A formatting model turns that draft into clean writing. Both run on your Mac, and ZenVoice downloads only the ones you pick.
Voice models
The model that listens. Start with a pick for what you need; switch any time in the app.
English or a European language
Parakeet Ultra
Our most accurate, in 25 European languages.
- Languages
- 25 European
- Typical delay
- 0.11 s per clip
- 7.4%
- WER
- 71×
- real time
- 0.9 GB
- download
Words as you speak
Voxtral Realtime
Writes as you speak, punctuation included.
- 9.5%
- WER
- 4×
- real time
- 2.8 GB
- download
Nothing to download
Apple Speech Analyzer
Built into macOS 26. Nothing to download.
- 10.0%
- WER
- 60×
- real time
- Built in
- size
Auto-detect, or one of 99 languages
Whisper Large V3 Turbo
Detects your language for you, across 99.
- 9.7%
- WER
- 10×
- real time
- 0.5 GB
- download
1,600+ languages
Omnilingual ASR
Tamil, Bengali, Swahili and 1,600 more languages.
- 15.6%
- WER
- 9×
- real time
- 1.0 GB
- download
The fastest English
NVIDIA Parakeet Unified
NVIDIA's newest English model, and the fastest.
- 8.1%
- WER
- 108×
- real time
- 0.7 GB
- download
Older or smaller Macs
Whisper Small
For Macs that can't run Turbo comfortably.
- 11.1%
- WER
- 32×
- real time
- 0.3 GB
- download
The largest model
Cohere Transcribe
Two billion parameters, the largest here.
- 8.3%
- WER
- 11×
- real time
- 2.9 GB
- download
All 13 voice models
The same columns for every model, so they compare at a glance.
ZenVoice Speech Bench V1 · MacBook Pro M5 · 24 GB
Apple Speech AnalyzerBuilt in
Apple · Built into macOS 26. Nothing to download.
System languages · 10.0% WER · 60× real time
System languages
10.0%
60×
Built in
Parakeet UltraRecommendedNew
Moondream · Our most accurate, in 25 European languages.
25 European · 7.4% WER · 71× real time
25 European
7.4%
71×
0.9 GB
Voxtral RealtimeStreamingNew
Mistral AI · Writes as you speak, punctuation included.
13 languages · 9.5% WER · 4× real time
13 languages
9.5%
4×
2.8 GB
Omnilingual ASRNew
Meta · Tamil, Bengali, Swahili and 1,600 more languages.
1,600+ languages · 15.6% WER · 9× real time
1,600+ languages
15.6%
9×
1.0 GB
NVIDIA Parakeet UnifiedNew
NVIDIA · NVIDIA's newest English model, and the fastest.
English · 8.1% WER · 108× real time
English
8.1%
108×
0.7 GB
NVIDIA Parakeet TDT 0.6B V3
NVIDIA · The model Parakeet Ultra builds on.
25 European · 7.7% WER · 70× real time
25 European
7.7%
70×
0.9 GB
Parakeet V3 · Neural EngineNew
NVIDIA ·
25 European
25 European
—
—
1.3 GB
NVIDIA Parakeet TDT 0.6B V2
NVIDIA · The English-only Parakeet.
English · 8.1% WER · 75× real time
English
8.1%
75×
0.8 GB
NVIDIA Nemotron 3.5 Multilingual 0.6BStreaming
NVIDIA · Streams as you speak in eight languages.
8 languages · 12.8% WER · 47× real time
8 languages
12.8%
47×
0.9 GB
Qwen3-ASR 0.6B
Qwen · A compact multilingual model, native on MLX.
Multilingual · 8.4% WER · 35× real time
Multilingual
8.4%
35×
0.8 GB
Cohere Transcribe
Cohere Labs · Two billion parameters, the largest here.
14 languages · 8.3% WER · 11× real time
14 languages
8.3%
11×
2.9 GB
Whisper Large V3 Turbo
OpenAI · Detects your language for you, across 99.
99 languages · 9.7% WER · 10× real time
99 languages
9.7%
10×
0.5 GB
Whisper Small
OpenAI · For Macs that can't run Turbo comfortably.
99 languages · 11.1% WER · 32× real time
99 languages
11.1%
32×
0.3 GB
Measured on 1,382 clips of real speech, every model on one Mac on October 27, 2026. Speed is times faster than real time. See the benchmarks →
Formatting models
Used by Smart formatting: fillers and restarts out; numbers, dates and money written the way you'd type them.
Apple Intelligence
Built into macOS on supported Macs. Nothing to download.
Built in
No download
Qwen3 1.7B
Helps Smart formatting split ambiguous spoken lists. English, offline.
0.98 GB
Download
Kanso Koi 1.0, Kanso Satori 1.0 and Kanso Koi 0.9 are retired from ZenVoice. Install them from Terminal →