Hold Fn, talk, get text worth keeping
A speech model on your Mac's Neural Engine transcribes while a local AI polishes the result: your jargon survives, your false starts don't, and ⌘Z always restores your exact words.
Claude CodeHow do I dictate text on a Mac and get clean writing out?
Raw speech-to-text types your false starts, filler and half-corrections verbatim, so dictation worth keeping is two stages: a transcription model turns audio into words, then a language model rewrites that transcript into the sentence you meant. Both stages can run on-device on Apple silicon. PortBay's dictation does exactly that behind a held Fn key, with each half independently swappable: transcription is macOS Dictation by default, or a local Whisper or Parakeet model on the Neural Engine (macOS 14+); the rewrite is Apple Intelligence by default on macOS 26+, or your local Ollama server. The rewrite is vocabulary-aware — custom terms, project names and jargon learned from past dictations steer corrections — and output that adds facts the transcript did not contain is rejected. ⌘Z always restores your exact words, the last 20 dictations are kept locally for recovery, and audio never leaves the machine.
Verified against the product, 2026-07-10.
From notch to commit message
Hold Fn in any app: the notch HUD shows the live waveform, the local model transcribes, and the polished text lands at your cursor.
From voice to clean text
- Hold Fn and talk
- The local model transcribes
- A local AI polishes it
- Undo is a contract
- Take it to every app
Hold Fn and talk
In any PortBay field out of the box: a card title, an agent prompt, a commit message. A quick tap stays whatever your Fn key already does; only a held press dictates.
The local model transcribes
Pick your engine: Parakeet TDT v3 (~2.4 GB, 25 European languages, fastest on Apple Silicon) or Whisper Large v3 Turbo (~1.6 GB, 99 languages). Streaming models show live captions while you speak.
whisper-large-v3-turbo · Neural EngineA local AI polishes it
A light touch-up when your speech was clean, a restructure when it rambled. Custom terms and on-screen vocabulary steer it ('shop if I' becomes Shopify); output that invents content is rejected.
Undo is a contract
⌘Z restores your exact words, and if the rewrite provider is unavailable the raw transcript stands. Dictation degrades to plain typing, never to lost text.
Take it to every app
Enable Dictate anywhere (a local model plus the macOS Accessibility permission). Hold Fn in your editor, browser or Slack; double-tap for hands-free; Esc cancels; the last 20 dictations are recoverable from the tray.
Dictation, tier by tier
Needs a downloaded local model and the Accessibility permission.
Comma-separated jargon list; the first 12 are applied when something like them was spoken.
Local engine needs macOS 14 or later; Apple Intelligence rewrites need macOS 26 with Apple Intelligence enabled. Everything here ships in the free tier.
The dictation stack, assembled from parts
The built-in option types your false starts verbatim; the DIY route gives you a transcript file and no rewrite, no vocabulary, no undo contract.
PortBay bundles the models, the rewrite, the vocabulary learning and the recovery net, in the tool your tasks already live in.
# macOS dictation: your ums and corrections, typed faithfully # Or wire whisper.cpp yourself$ git clone https://github.com/ggml-org/whisper.cpp && cd whisper.cpp$ make && ./models/download-ggml-model.sh large-v3-turbo$ ./main -m models/ggml-large-v3-turbo.bin -f recording.wav# record the wav yourself, paste the output yourself# no jargon handling, no rewrite, no notch HUD, no ⌘Z contract # Or a separate paid dictation app, one more subscription
One app, not a stack of tools
Local AI
A managed, supervised Ollama server with a live model catalog and one on-device endpoint shared by dictation, agents and your own tools.
Explore →Agent task board
A Kanban board both you and your agents work. Cards are Markdown in the repo; moving one to To Do dispatches the agent.
Explore →Agent dev environment
Stack detection, runtimes, HTTPS, databases and mail, provisioned per project — the running app your agents verify their work against.
Explore →Asked before downloading
No. The macOS engine captures audio in the system's own corespeechd; the local engine captures in PortBay's bundled speech sidecar and transcribes on the Neural Engine. Only the text transcript goes to the rewrite model, and both providers (Apple Intelligence, your Ollama server) run on the machine.
Parakeet TDT v3 (~2.4 GB, 25 European languages, near-instant on Apple Silicon), Whisper Large v3 Turbo (~1.6 GB, 99 languages, the default Whisper pick), Distil-Whisper Large v3 and Whisper Medium for English, and Whisper Large v3 when accuracy beats latency. Managed from the AI page, stored beside your Ollama models.
Yes. With a local model downloaded and the macOS Accessibility permission granted, hold Fn in any app and the transcript is typed at your cursor. The HUD lives in the camera notch (or under the menu bar), double-tap Fn goes hands-free, and Esc cancels without typing anything.
⌘Z restores your exact words, always. The rewrite also rejects its own output when it adds information the transcript did not contain, and if the provider is unavailable your words stay as transcribed. The failure mode is plain dictation, never lost text.
Yes. Select text, hold Fn, and speak an instruction: 'make this a bullet list', 'tighten this paragraph'. The selection is transformed in place instead of appended to.
Give your projects and your agents a real local home.
Download for macOSFree & open source · macOS 11+ on Apple Silicon · Pro from $10/mo