Whisper dictation on the Mac: the model is the easy part.
Getting Whisper to transcribe a file takes an afternoon. Getting it to feel like a keyboard takes a hotkey daemon, an insertion path that survives Electron apps, and a dictionary that learns. That is the gap.
Whisper is free and whisper.cpp runs it fast on Apple Silicon, so plenty of people start by scripting it themselves. That works until you want push-to-talk from any app, text that lands in the right field, punctuation that survives, and a fix for the three words it always gets wrong. Those are the parts a dictation app is actually selling.
What sits between a model and a keyboard.
Capture that starts and stops cleanly
Blurt uses a push-to-talk key, Fn by default, with Right Command, Right Option or Right Control as alternatives and a configurable fallback shortcut. Hold to talk, tap to toggle, or live mode where text streams as you speak.
Insertion that works everywhere
Text is inserted through a paste with full clipboard restore, so Electron apps, browsers and Terminal all behave. It needs the Accessibility permission. If insertion fails, the text stays in history and the heads-up display tells you rather than dropping it.
Accuracy you can steer
Custom vocabulary biases the model toward your product names. Replacement rules fix words it consistently mishears. Smart match catches near misses, and Blurt proposes corrections it notices in your own history. A raw whisper.cpp pipeline has none of this.
A second engine when Whisper is not the answer
Parakeet TDT 0.6B v3 runs on the CPU, covers English and 25 European languages, and returns text faster than Whisper. You switch in settings and the model downloads in place.
| Blurt | whisper.cpp by hand | |
|---|---|---|
| Cost | $3.99/mo, $29/yr, or $99 once | Free, plus your time |
| Push-to-talk from any app | Built in, four key choices, three modes | You write the daemon |
| Inserting into the focused app | Paste with clipboard restore, works in Electron and Terminal | Scriptable, fragile across apps |
| Model management | Pick a quality tier, it downloads in place | Manual downloads and flags |
| Vocabulary and corrections | Vocabulary, rules, smart match, learns from edits | Prompt hints only |
| Formatting | Local model for punctuation, paragraphs and lists | Whatever you build |
| History | Searchable, editable, exportable | Your log files |
| Updates | Signed and notarised, verified with minisign | git pull |
Common questions
Which Whisper model does Blurt use?
Whisper large-v3-turbo, in three quantised builds: 574 MB compact, 874 MB accurate as the default, and 1.6 GB for maximum accuracy. All run on the Metal GPU. Parakeet is offered as a faster CPU alternative for English and 25 European languages.
Is it worth paying when Whisper is free?
If you enjoy maintaining a hotkey daemon and an insertion script, no. If you want dictation that works in every app tomorrow morning and keeps improving on your vocabulary, that is what the price buys.
Can I bring my own model?
You choose between the bundled Whisper quality tiers and Parakeet in settings. Arbitrary custom weights are not supported. If loading your own fine-tune matters, an open source app like VoiceInk is a better fit.
Does any of it touch the cloud?
No. Models come from Hugging Face on first download, and after that transcription and formatting both run locally. There is no API key field because there is nothing to call.
What are the requirements?
An Apple Silicon Mac, macOS 10.15 or later, the Accessibility permission, and enough disk for one model. There is no Intel build.
Whisper, wired to a key, in five minutes.
Download, pick a model, hold Fn and talk. Fourteen days unrestricted, no card and no account.
Apple Silicon. 14-day trial, no card and no account. After the trial it keeps working at 20 dictations a day. Every paid plan, Lifetime included, carries a 30-day money-back guarantee.