Free. Local. Powerful.

Press a key. Speak. Finished text appears at your cursor.

Download How it works

Speech to text runs offline, on your computer. For Windows and Macs with Apple Silicon, and Linux in alpha.

Ready Ctrl + Space Copy last Start meeting Open app Hide
Reply to Sam

Pressing the shortcut records, and the cleaned sentence is pasted into the window that had focus.

How dictation works

You control how much cleanup is applied to your voice typing.

Remove ‘um’ and ‘uh’, correct supported spellings and vocabulary, and format without rewriting.

SoIthinkweshouldshipthebetaonFriday.Let'skeeptheoldinstallerarounduntilwe'resure.

Whichever you pick, the raw recognition is saved unchanged alongside it.

Meeting transcription.

Your microphone and your computer’s audio, captured together and transcribed.

The Meetings page listing transcribed recordings with their dates and durations.
  • Notes are assembled from sentences that were actually said.
  • Import a recording you already have and transcribe it.
  • Export to HTML, PDF, Markdown or plain text.

Local and private.

Everything is processed on your own computer.

  • Recognition and cleanup run offline.
  • No account. Nothing is uploaded unless you send a report.
  • An outside model is used only if you connect one yourself.
The Dictation page showing the text cleanup choice and the local cleanup model.

Why these models.

We compared open speech models by download size and published accuracy before choosing which ones SonoJot offers.

Qwen makes fewer English word errors than Whisper at a similar size, which is why SonoJot recommends it.

Offered in SonoJot Left out Best accuracy for its size
Models that run on a CPU
Word errors (%), lower is better 6 7 8 9 0.10.31310 Download GB (log scale) Qwen 1.7B Q4: 1.64 GB, 6.25% word errors. Offered. Our most accurate measured CPU dictation option. Qwen 0.6B Q8: 1.02 GB, 6.93% word errors. Offered. Smaller multilingual option for less powerful computers. Qwen 1.7B sherpa: 2.40 GB, 6.25% word errors. Left out. Larger and slower than Qwen Q4 in SonoJot's tests. Parakeet v3: 0.67 GB, 6.82% word errors. Left out. More holdout errors than v2 in our sherpa CPU full-clip comparison. Parakeet v2: 0.66 GB, 6.44% word errors. Offered. Compact English speech recognition on CPU. Parakeet 110M: 0.14 GB, 7.98% word errors. Left out. Not yet tested through SonoJot's recognition path. Canary 180M: 0.21 GB, 7.48% word errors. Left out. Needs a speech adapter SonoJot doesn't ship. Canary 1B Flash: 1.05 GB, 6.81% word errors. Left out. Needs a speech adapter SonoJot doesn't ship. Moonshine Medium: 0.27 GB, 7.18% word errors. Offered. Fast English streaming with live text. Moonshine Small: 0.14 GB, 8.42% word errors. Offered. Small English streaming option. Whisper Turbo: 1.04 GB, 8.44% word errors. Offered. Kept for languages Qwen does not cover; GPU support requires Apple Metal. Whisper Medium.en: 0.95 GB, 8.66% word errors. Left out. More errors than smaller CPU options. Whisper Small.en: 0.38 GB, 9.24% word errors. Offered. Multilingual sibling kept for languages Qwen lacks; plotted WER is Small.en.
  1. 1Parakeet v3
  2. 2Canary 1B Flash
Word errors (%), lower is better 6 7 8 9 0.10.31310 Download GB (log scale) Qwen 1.7B Q4: 1.64 GB, 6.25% word errors. Offered. Our most accurate measured CPU dictation option. Qwen 0.6B Q8: 1.02 GB, 6.93% word errors. Offered. Smaller multilingual option for less powerful computers. Qwen 1.7B sherpa: 2.40 GB, 6.25% word errors. Left out. Larger and slower than Qwen Q4 in SonoJot's tests. Parakeet v3: 0.67 GB, 6.82% word errors. Left out. More holdout errors than v2 in our sherpa CPU full-clip comparison. Parakeet v2: 0.66 GB, 6.44% word errors. Offered. Compact English speech recognition on CPU. Parakeet 110M: 0.14 GB, 7.98% word errors. Left out. Not yet tested through SonoJot's recognition path. Canary 180M: 0.21 GB, 7.48% word errors. Left out. Needs a speech adapter SonoJot doesn't ship. Canary 1B Flash: 1.05 GB, 6.81% word errors. Left out. Needs a speech adapter SonoJot doesn't ship. Moonshine Medium: 0.27 GB, 7.18% word errors. Offered. Fast English streaming with live text. Moonshine Small: 0.14 GB, 8.42% word errors. Offered. Small English streaming option. Whisper Turbo: 1.04 GB, 8.44% word errors. Offered. Kept for languages Qwen does not cover; GPU support requires Apple Metal. Whisper Medium.en: 0.95 GB, 8.66% word errors. Left out. More errors than smaller CPU options. Whisper Small.en: 0.38 GB, 9.24% word errors. Offered. Multilingual sibling kept for languages Qwen lacks; plotted WER is Small.en.
  1. 1Parakeet 110M
  2. 2Canary 180M
  3. 3Parakeet v3
  4. 4Canary 1B Flash
  5. 5Qwen 1.7B sherpa
Models that use a GPU
Word errors (%), lower is better 6 7 8 9 0.10.31310 Download GB (log scale) Qwen 1.7B Q8: 2.52 GB, 6.25% word errors. Offered. Our GPU dictation choice, informed by tests on real dictation. Qwen 1.7B FP16: 4.73 GB, 6.25% word errors. Left out. Larger download and slower response than Qwen Q8. Granite 4.1 NAR: 2.50 GB, 5.70% word errors. Offered. For meetings: fewer errors on room microphones in our tests. Granite 4.1 AR: 2.56 GB, 5.65% word errors. Left out. Similar accuracy to NAR, much slower in our measured runtime. Granite 4.0: 2.56 GB, 5.87% word errors. Left out. Superseded by 4.1; repeated output on some test clips. Granite 3.3 2B: 3.20 GB, 6.35% word errors. Left out. Needs a runtime SonoJot doesn't ship. Granite 3.3 8B: 9.50 GB, 6.07% word errors. Left out. Much larger download for a small accuracy gain. Cohere Transcribe: 2.41 GB, 5.84% word errors. Left out. Did not improve dictation over Qwen in our tests. Canary-Qwen: 2.80 GB, 6.05% word errors. Left out. More errors and slower than Qwen on our test recordings. Parakeet v3: 0.67 GB, 6.82% word errors. Left out. More holdout errors than v2 in our sherpa CPU full-clip comparison. Canary 1B Flash: 1.05 GB, 6.81% word errors. Left out. Needs a speech adapter SonoJot doesn't ship. Whisper Turbo: 1.04 GB, 8.44% word errors. Offered. Kept for languages Qwen does not cover; GPU support requires Apple Metal. Whisper Large v3: 1.78 GB, 7.95% word errors. Left out. More word errors than Qwen; slow on CPU. Voxtral Mini 3B: 3.19 GB, 7.53% word errors. Left out. Less accurate than Qwen at a larger size. Voxtral Realtime: 2.60 GB, 8.24% word errors. Left out. More errors than Qwen on the published English average. Voxtral Small 24B: 14.00 GB, 7.06% word errors. Left out. Much larger than Qwen without an accuracy gain. Phi-4 multimodal: 4.00 GB, 6.46% word errors. Left out. Needs a runtime SonoJot doesn't ship.
  1. 1Granite 4.1 AR
  2. 2Granite 4.0
  3. 3Canary-Qwen
  4. 4Granite 3.3 2B
  5. 5Qwen 1.7B FP16
  6. 6Voxtral Small 24B
Word errors (%), lower is better 6 7 8 9 0.10.31310 Download GB (log scale) Qwen 1.7B Q8: 2.52 GB, 6.25% word errors. Offered. Our GPU dictation choice, informed by tests on real dictation. Qwen 1.7B FP16: 4.73 GB, 6.25% word errors. Left out. Larger download and slower response than Qwen Q8. Granite 4.1 NAR: 2.50 GB, 5.70% word errors. Offered. For meetings: fewer errors on room microphones in our tests. Granite 4.1 AR: 2.56 GB, 5.65% word errors. Left out. Similar accuracy to NAR, much slower in our measured runtime. Granite 4.0: 2.56 GB, 5.87% word errors. Left out. Superseded by 4.1; repeated output on some test clips. Granite 3.3 2B: 3.20 GB, 6.35% word errors. Left out. Needs a runtime SonoJot doesn't ship. Granite 3.3 8B: 9.50 GB, 6.07% word errors. Left out. Much larger download for a small accuracy gain. Cohere Transcribe: 2.41 GB, 5.84% word errors. Left out. Did not improve dictation over Qwen in our tests. Canary-Qwen: 2.80 GB, 6.05% word errors. Left out. More errors and slower than Qwen on our test recordings. Parakeet v3: 0.67 GB, 6.82% word errors. Left out. More holdout errors than v2 in our sherpa CPU full-clip comparison. Canary 1B Flash: 1.05 GB, 6.81% word errors. Left out. Needs a speech adapter SonoJot doesn't ship. Whisper Turbo: 1.04 GB, 8.44% word errors. Offered. Kept for languages Qwen does not cover; GPU support requires Apple Metal. Whisper Large v3: 1.78 GB, 7.95% word errors. Left out. More word errors than Qwen; slow on CPU. Voxtral Mini 3B: 3.19 GB, 7.53% word errors. Left out. Less accurate than Qwen at a larger size. Voxtral Realtime: 2.60 GB, 8.24% word errors. Left out. More errors than Qwen on the published English average. Voxtral Small 24B: 14.00 GB, 7.06% word errors. Left out. Much larger than Qwen without an accuracy gain. Phi-4 multimodal: 4.00 GB, 6.46% word errors. Left out. Needs a runtime SonoJot doesn't ship.
  1. 1Parakeet v3
  2. 2Canary 1B Flash
  3. 3Cohere Transcribe
  4. 4Granite 4.1 AR
  5. 5Granite 4.0
  6. 6Voxtral Realtime
  7. 7Canary-Qwen
  8. 8Voxtral Mini 3B
  9. 9Granite 3.3 2B
  10. 10Phi-4 multimodal
  11. 11Qwen 1.7B FP16
  12. 12Granite 3.3 8B
  13. 13Voxtral Small 24B

Word errors are the Open ASR Leaderboard's published English average across seven public test sets, excluding TED-LIUM, from the snapshot of October 12, 2026 (revision f82da858). They are scores for each model family, not measurements of the quantized files SonoJot downloads, and they say nothing about speed or other languages. Error rates do not depend on hardware; the two views only separate models by where they run.

Sizes exclude runtimes, and some sizes for models SonoJot does not ship are estimates. Whisper Small and Medium use their English-only sibling's scores. Models without a comparable score, including Apple Speech, are not shown. SonoJot's own tests on real dictation also shaped which models it offers.

Leaderboard data
Model table
ModelDownloadWord errorsRuns onIn SonoJotWhy
Granite 4.1 NAR 2.50 GB 5.70% GPU Offered For meetings: fewer errors on room microphones in our tests.
Qwen 1.7B Q8 2.52 GB 6.25% GPU Offered Our GPU dictation choice, informed by tests on real dictation.
Qwen 1.7B Q4 1.64 GB 6.25% CPU Offered Our most accurate measured CPU dictation option.
Parakeet v2 0.66 GB 6.44% CPU Offered Compact English speech recognition on CPU.
Qwen 0.6B Q8 1.02 GB 6.93% CPU Offered Smaller multilingual option for less powerful computers.
Moonshine Medium 0.27 GB 7.18% CPU Offered Fast English streaming with live text.
Moonshine Small 0.14 GB 8.42% CPU Offered Small English streaming option.
Whisper Turbo 1.04 GB 8.44% CPU, GPU Offered Kept for languages Qwen does not cover; GPU support requires Apple Metal.
Whisper Small.en 0.38 GB 9.24% CPU Offered Multilingual sibling kept for languages Qwen lacks; plotted WER is Small.en.
Granite 4.1 AR 2.56 GB 5.65% GPU Left out Similar accuracy to NAR, much slower in our measured runtime.
Cohere Transcribe 2.41 GB 5.84% GPU Left out Did not improve dictation over Qwen in our tests.
Granite 4.0 2.56 GB 5.87% GPU Left out Superseded by 4.1; repeated output on some test clips.
Canary-Qwen 2.80 GB 6.05% GPU Left out More errors and slower than Qwen on our test recordings.
Granite 3.3 8B 9.50 GB 6.07% GPU Left out Much larger download for a small accuracy gain.
Qwen 1.7B sherpa 2.40 GB 6.25% CPU Left out Larger and slower than Qwen Q4 in SonoJot's tests.
Qwen 1.7B FP16 4.73 GB 6.25% GPU Left out Larger download and slower response than Qwen Q8.
Granite 3.3 2B 3.20 GB 6.35% GPU Left out Needs a runtime SonoJot doesn't ship.
Phi-4 multimodal 4.00 GB 6.46% GPU Left out Needs a runtime SonoJot doesn't ship.
Canary 1B Flash 1.05 GB 6.81% CPU, GPU Left out Needs a speech adapter SonoJot doesn't ship.
Parakeet v3 0.67 GB 6.82% CPU, GPU Left out More holdout errors than v2 in our sherpa CPU full-clip comparison.
Voxtral Small 24B 14.00 GB 7.06% GPU Left out Much larger than Qwen without an accuracy gain.
Canary 180M 0.21 GB 7.48% CPU Left out Needs a speech adapter SonoJot doesn't ship.
Voxtral Mini 3B 3.19 GB 7.53% GPU Left out Less accurate than Qwen at a larger size.
Whisper Large v3 1.78 GB 7.95% GPU Left out More word errors than Qwen; slow on CPU.
Parakeet 110M 0.14 GB 7.98% CPU Left out Not yet tested through SonoJot's recognition path.
Voxtral Realtime 2.60 GB 8.24% GPU Left out More errors than Qwen on the published English average.
Whisper Medium.en 0.95 GB 8.66% CPU Left out More errors than smaller CPU options.

The SonoJot punctuation model.

Speech recognizers put a period at many pauses. A small model of our own fixes sentence breaks and commas on your computer.

  • It places periods, commas, question marks and semicolons. It never adds, removes or changes a word.
  • It also marks spoken numbers, money and dates, so SonoJot can write “four thousand dollars” as $4,000.
  • 268 MB. About 0.17 seconds per 300 words on a desktop CPU once loaded.
Recognizer output before and after the model, on 60 real dictations held out from training (24,659 words)
BeforeAfter
Sentence breaks (F1)0.7470.886
False sentence breaks per 1,000 words16.70.9
Commas (F1)0.7440.884

Fine-tuned from Horizon-Labs punctuation-restoration-small (mmBERT-small) on earnings-call transcripts, synthetic dictation-style text and one consenting user's dictation. English only. Capitals come from SonoJot's rules, not the model. Open weights under CC BY-SA 4.0.

SonoJot Punctuation on Hugging Face

Install.

Download the SonoJot dictation app for Windows or macOS on Apple Silicon, or build the Linux alpha.

Windows

beta

Windows 10 and 11, 64-bit

Download
PowerShell 
irm https://sonojot.com/install.ps1 | iex

macOS

alpha

Apple Silicon, macOS 14 and later

Download
Terminal 
curl --proto '=https' --proto-redir '=https' -fsSL https://sonojot.com/install.sh | sh

Linux

alpha

Arch Linux and Omarchy, x86_64; Hyprland for the shortcut and paste

Early and rough: expect bugs and missing pieces. It builds from source, and meetings record the microphone only. Use Feedback in the app to report problems.

Build from source
Terminal 
git clone https://github.com/Cody-Stevens/sonojot.git
sonojot/rewrite/scripts/linux/install.sh

Windows may show a SmartScreen prompt because the installer is not Authenticode-signed yet; choose More info → Run anyway.

Disclaimer: I made SonoJot for myself and use it daily. It is free and open source under the Apache License 2.0: fork it, change it, keep it. It comes with no warranty. Reported problems may or may not be addressed, and the app will change as I see fit, without notice.