Local Whisper on a Mac
You want the words, and you do not want the recording to leave the computer to get them.
Short answer: Local Whisper is OpenAI's speech model running on your Mac through whisper.cpp, so dictation audio is not uploaded for transcription. MightyMouse tries NVIDIA Parakeet TDT 0.6B v3 first for the 25 European languages that model covers, including English, and falls back to Whisper when Parakeet is missing or the language is outside that set.
Whisper is a speech model you can run on your own computer. "Local" means the audio file is transcribed by a program on the Mac, not by sending the clip to someone else's server. That is the whole reason people search for it.
How to dictate on a Mac with local Whisper
- Click into the app where you want the text.
- Press the listen key and talk. Press it again to stop. You do not hold the key.
- The Mac transcribes the clip. For a covered language, NVIDIA Parakeet TDT 0.6B v3 runs first if it is installed. Otherwise Whisper runs.
- The text is typed at the cursor.
English Whisper uses the base model. The file on disk is ggml-base.en.bin. Any other language uses the multilingual Whisper build, ggml-large-v3-turbo.bin, because the English-only file cannot decode it.
NVIDIA Parakeet TDT 0.6B v3 is used under CC-BY-4.0. The engine around it is NeMo-Speech, Apache-2.0. Naming the model here is the attribution that license asks for.
What stays on the Mac, and what does not
Transcription does not upload the audio. If you then ask a cloud model to rewrite the text, that text leaves the machine, the same as if you had pasted it into the model's own app. The dictation step and the "now do something with it" step are separate, which is the point of dictating anywhere on your Mac.
It also works with no internet, for the transcription itself. A cloud dictation app cannot say that, because the clip has to reach their server before you get words.
What local transcription costs you
- Cleanup is weaker. A large cloud model is better at tidying a long ramble. Local models are good at the words and do less rewriting.
- It uses your Mac. The model runs on your CPU or Neural Engine while you talk.
- It is not a file tool. This is live dictation at the cursor. A recording you already have is MacWhisper's job.
There is no word cap and no per-minute meter on the transcription, because there is no server to meter. Wispr Flow is the comparison if you want the cloud cleanup and you are fine sending the audio. Apple Dictation can also run on device on Apple silicon, with a 30 second silence limit and no custom vocabulary. The roundup is the best dictation app for Mac.
Hands-free, and the terminal
The same on-device path is what hands-free dictation uses once a session is running: you talk, you pause, the sentence is typed, you tap to stop. Over SSH, the thing that matters is not which model ran. It is whether the text arrives as keystrokes. That failure mode is why dictation stops working in Terminal and over SSH.
Talk instead of typing. The transcription stays on your Mac.
No word cap. No card to try it.
Common questions
Does local Whisper upload my voice?
No. The audio is transcribed on your Mac. If you later send the text to a cloud model, that text leaves, the same as pasting it yourself. The recording does not go out for the transcription step.
Which model does MightyMouse actually run?
If NVIDIA Parakeet TDT 0.6B v3 is installed and your language is one of the 25 European languages it covers, including English, that runs first. Otherwise Whisper runs through whisper.cpp. English Whisper uses the base model file ggml-base.en.bin unless you switch it.
Does it work offline?
Yes, once the model files are on the Mac. Transcription does not need a network. Asking a cloud model to edit the text afterwards does.
Is local Whisper as polished as a cloud dictation app?
It is accurate at the words and weaker at rewriting a long ramble. If you want that cleanup, a cloud tool such as Wispr Flow does it by sending the audio to their servers.
If you want this on your own machine: MightyMouse dictates into any Mac app with transcription that stays on your Mac, and the same phone can be the mic.