It all runs locally
Speech is recognised on your own processor. The network is needed exactly once - to fetch the model.
Dictation for macOS
Speech into text, entirely on your Mac.
Hold a key, talk, let go - the finished text lands wherever the cursor is. Recognition runs on the machine itself, so no audio goes anywhere. No account, no subscription, no cloud.
Requires macOS 26 or later.
Didn't install, or isn't working? Message me on Telegram or at , and I'll sort it out.
Why bother
Four reasons, not a checklist
Nothing is uploaded
Not the audio, not the transcripts. There is no server and no account, so there is nothing to leak.
Intel Macs count
Metal where it exists, the CPU where it doesn't. Apple Silicon is not a requirement.
Free
A non-commercial project. No tiers, no subscription, no cap on minutes.
How it works
Four steps, all inside one keypress
The sequence is the same in every app: mail, a code editor, a chat window, a search field.
⌥
Hold the key
Right ⌥ by default. Settings can switch it to fn or Right ⌘.
Speak
Text starts right away
The model streams: words show up in the floating capsule while you are still talking.
Release
Recording stops
Let the key go and the recording ends, with the transcript finished off.
▏
The text is there
Cleaned-up text is inserted into the focused field. A copy stays in the history.
The engine
Nemotron 3.5 ASR Streaming, 0.6B parameters
The model runs locally through transcribe.cpp on top of ggml - Metal when it is available, CPU otherwise. That is exactly why the app works on Intel Macs.
- Model
- Nemotron 3.5 ASR Streaming 0.6B, ~716 MB Downloaded on demand from inside the app and checked against a published SHA256.
- Languages
- 40 locales Russian and English are both in the top, transcription-ready tier.
- Second engine
- Apple SpeechAnalyzer - needs Apple Silicon On Intel Macs macOS reports it as unavailable for every language. The app says so plainly instead of failing silently.
- Licences
- ggml is MIT; the model weights are OpenMDW-1.1
MEASURED ON AN INTEL MAC 4.3 s of speech → 1.3 s to transcribe ≈3.3× realtime
Inside the app
What separates dictation from transcription
Names and jargon the engine keeps mishearing
Two kinds of rule: corrections - when you hear X, write Y - and terms that bias the recognition itself in the right direction.
No filler words, no stray spaces
A deterministic pass strips filler and fixes spacing and punctuation. Optionally a smarter pass runs through an on-device Apple Intelligence model, on Macs that support it.
Everything you dictated stays to hand
A searchable list of past transcriptions; any of them copies with a single click.
Russian and English
The interface language switches in Settings and applies instantly, with no restart.
The one you want, not the one you got
With several microphones around - built-in, USB, a headset - VSTT records through whichever one the system has chosen, and follows it when you change that. Want a particular one? Pick it in Settings under "Microphone"; the device itself is remembered rather than its name, so renaming it, or owning a second identical one, changes nothing. If the one you picked isn't connected, dictation goes through the system's and the capsule says so.
Lives in the tray, not the Dock
After launch it's just a menu bar icon - no Dock icon. "Open VSTT" there shows the window; close it - ⌘W or the red button - and it settles back into the tray. The dictation key works either way.
Interface
What it looks like
Requirements
What you need to run it
- System
- macOS 26 or later
- Processor
- Apple Silicon or Intel Nemotron runs on both. Apple SpeechAnalyzer is Apple Silicon only.
- Disk space
- ~716 MB for the model, plus the app itself (5.8 MB)
- Network
- Only for the first model download After that the app works with no internet at all.
- Smart cleanup
- Optional - needs a Mac that supports Apple Intelligence
Try it on one sentence
Download it, let the app fetch the model, then hold Right ⌥.
Requires macOS 26 or later.