Instant Mac dictation · Open source
Faster and more accurate than built-in dictation — in any app. Teach it your vocabulary or use a compatible fine-tuned model.
Prompt
Describe the change…
No account and no upload. Downloads happen once; the model runs on your own silicon.
Others pick one language and translate the rest away. This one keeps them all.
Twelve seconds of speech, transcribed in under one second.
Add a name, term, or jargon once. It remembers from then on.
Voice shortcuts
Turn a short spoken trigger into a URL, signature, address, reusable reply, or any text you type often. Shortcuts stay on your Mac and work wherever the cursor is.
shortcut github
https://github.com/yourname
Saved locally · available in every text field
The idea
Chunked transcription slices audio into pieces, and the cuts land mid-word — exactly where a mixed sentence breaks. WhisprStream re-reads the whole utterance instead, several times a second.
Chunked · the cut lands mid-word
Whole · read again from the start
Bring the right model
Use another supported transcription model or a compatible fine-tuned checkpoint. Even when the model only returns a finished transcript, WhisprStream repeatedly re-reads the growing utterance and settles the stable words into a live stream.
Parameters. Small enough to stay resident and warm.
To transcribe twelve seconds of bilingual speech.
Leave the machine. There is no server to leave for.
MIT licensed
Software that listens to you should be software you can audit.
↓Get the app ZIP from Releases
View on GitHub