OpenAI’s Whisper is an excellent speech-to-text model, and Apple Silicon runs it fast. StarWhisper packages Whisper into a native macOS app so you can use it for real-time dictation without building anything.
Whisper is open source, and the whisper.cpp project made it run efficiently on Apple Silicon using Metal. The catch is setup: compiling, downloading model files, and wiring up audio. StarWhisper does all of that for you in a signed, notarized Mac app.
Download the DMG, open the app, and dictate. Models download inside the app, Metal acceleration is automatic, and text flows into any app you are using.
Clone the repo, build with Metal enabled, fetch a model, and run it from the terminal per file. Great for scripting, slower for everyday dictation.
Whisper comes in several sizes. Bigger models are more accurate but need more memory and time:
StarWhisper lets you switch models in settings, and uses a lighter model for the real-time preview while a larger one produces the final text.
The unified memory and GPU on M1, M2, M3, and M4 chips let Whisper run with Metal acceleration entirely on-device. That means low latency for live dictation and full privacy, since the audio never leaves your Mac.
Yes, and it runs well. StarWhisper runs Whisper locally with Metal GPU acceleration on M1, M2, M3, and M4 Macs.
No. StarWhisper is a normal Mac app. If you prefer the manual route, whisper.cpp is open source and can be built from source, but it is not required.
Yes. StarWhisper uses Apple’s Metal framework to run Whisper on the GPU for faster transcription.
Small is a good balance for live dictation. Use large-v3 for maximum accuracy, or tiny and base for the fastest quick notes.
No. Transcription happens on your Mac. There is no cloud step and no audio stored on a server.
Skip the build steps. Download the app and start dictating with Whisper in minutes.
Download for MacFree to start. Works on macOS 12 or later, Apple Silicon only.