The problem
Most dictation tools upload audio to a service to transcribe it. If the material is under NDA, covered by a contract, or a clinical note, that upload is the problem, and a privacy policy does not undo it.
Running speech models locally on Apple Silicon is possible. The work is in the plumbing: fetching and storing model weights, wiring Metal and CoreML, resampling audio to what the model expects, and handling the machine going to sleep mid-recording.
Raw speech-to-text is not usable text. It arrives without punctuation or casing, and sending it to a cloud model to fix that hands back exactly the privacy the local transcription just bought.
Approach
Tract is a Tauri 2 application: a Rust core, a React interface, and a Swift bridge for the macOS APIs with no Rust binding — App Intents, ScreenCaptureKit, the Speech framework.
Transcription runs through whisper.cpp with Metal and CoreML switched on. The default model is Whisper large-v3-turbo at roughly 1.55 GB on disk. Smaller weights are selectable down to 80 MB, and a quantised large-v3-turbo sits in between at 600 MB.
Parakeet TDT 0.6B, about 460 MB, is the second engine. It runs int8 ONNX on the CPU through sherpa-onnx, which matters on an Intel Mac or when the GPU is busy.
The cleanup pass — punctuation, casing, the obvious misheard word — uses Llama 3.2 3B Instruct at 4-bit through Apple's MLX. That stays on the machine too, because moving it off would defeat the point.
How it works
Audio comes off the selected input and is resampled to 16 kHz mono, which is what both engines expect.
The selected model runs on the machine. whisper.cpp uses Metal with flash attention; Parakeet runs on the CPU. Weights are already on disk, so no request goes out while you are talking.
A local Llama 3.2 3B pass adds punctuation and casing and fixes obvious errors, turning a transcript into something you can paste.
The text goes where you are working. An optional encrypted sync keeps a history across your own machines.
Features
Fit
Questions
The things people check before asking for a number.
No. Both engines run locally against weights already on disk, and the cleanup pass runs locally too. The network is used to download a model the first time you select it, and by two optional features you can leave switched off: encrypted sync, which encrypts on your machine before upload, and a cloud rewrite that needs your own API key.
Whisper large-v3-turbo is the default and the most accurate, at about 1.55 GB. The quantised build at 600 MB is the usual trade if disk or memory is tight. Parakeet TDT at roughly 460 MB runs on the CPU, which is the one to pick on an Intel Mac.
No, and there is no near-term plan for one. The application depends on macOS frameworks through a Swift bridge, so a port would mean replacing that layer rather than recompiling.
It is at 0.1.0 and in daily use by a small group. It builds signed universal installers, but auto-update is off, crash reporting is local only, and some icons are placeholders. Treat it as a working tool that is still moving.
Access is limited during the private beta. Check https://tract.nexg.dev for current status, or contact NexG.
Get a quotation
Share the workflow, users, and constraints you're working with. We'll follow up with a scoped quotation.