okoflow 开源 macOS 听写工具 Dictate，本地 Whisper 加云端改写

Dictate

Push-to-talk dictation for macOS with on-device Whisper.

dictate.okoflow.com

Hold the right Option key, speak, and let go: the text appears wherever you are typing, in any app. Whisper turns your voice into text on your Mac, so the audio never leaves it. When you want polished prose rather than a transcript, a cloud mode rewrites the text with Claude or OpenAI.

What it does

• Types into any app. The text is pasted into the focused field, and your clipboard is put back afterward. Password fields are left alone.
• Recognizes speech on your Mac. Whisper large-v3 turbo runs on Core ML through WhisperKit. Audio stays in memory and is never saved.
• Understands thirteen languages. Chinese, Dutch, English, French, German, Italian, Japanese, Korean, Polish, Portuguese, Russian, Spanish, and Ukrainian. Dictate detects which one you speak among those you choose.
• Cleans up as much as you like. Light removes hesitations such as "um", capitalizes the first letter, and adds a full stop, offline. Clean, Formal, and Translate to English rewrite the text with Claude or OpenAI, whichever you choose. Any app can have a mode of its own.
• Learns your words. The dictionary fixes names and terms that come out wrong, and snippets turn a spoken phrase into text such as your email address.
• Remembers when you want it to. The last 50 dictations stay on your Mac, searchable and one click away from the clipboard.

Install

Dictate is built from source. You need a Mac with Apple silicon, macOS 14 or later, and the Command Line Tools for Xcode 26 or later; a full Xcode works too.

git clone https://github.com/okoflow/dictate.git
cd dictate
make signing
make run

make signing runs once: it creates a local code-signing identity, so macOS keeps the permissions you grant across rebuilds, and asks for your password to trust it. make run builds build/Dictate.app, signs it, and opens it. Copy the app to /Applications to keep it.

On first launch, a short guide asks for the microphone and Accessibility permissions while Dictate downloads the speech model, about 630 MB. Loading the model for the first time compiles it for your chip and takes about a minute; after that it takes a second.

Use it

1. Click into a text field in any app.
2. Hold right Option and speak.
3. Let go. The text appears a moment later.

Dictate lives in the menu bar. Its menu switches the mode and the language, copies the last transcript, and opens Settings, where you can choose right Command or right Shift as the dictation key, pick a microphone, and choose the languages Dictate listens for.

Modes

Light is the default. Clean, Formal, and Translate to English need an API key from Anthropic or OpenAI, which you add in Settings › Modes and which stays in your Keychain. Without a key, offline, or when the provider does not answer within 3 seconds, Dictate uses Light and says why, so you always get your text.

Privacy

Speech recognition, Raw, and Light run entirely on your Mac. Dictate uses the network for two things only: downloading the speech model once, and the cloud modes, which send the recognized text, never audio, to Anthropic or OpenAI, whichever you choose. Dictated text never reaches the logs. Privacy lists every file Dictate keeps and how to remove it.

Documentation

• Usage: modes, languages, the dictionary, snippets, history, and settings.
• How it works: what happens between the key press and the pasted text.
• Privacy: what stays on your Mac and what leaves it.
• Troubleshooting: permissions, the dictation key, the speech model, and logs.
• Architecture: the modules, their interfaces, and how they fit together.

Contributing

CONTRIBUTING.md covers the development setup, the checks that run in CI, and the commit and pull request conventions.

Security

Report vulnerabilities privately through GitHub security advisories. SECURITY.md describes the process and the scope.

License

Copyright The Dictate Authors, listed in AUTHORS.

MIT. See LICENSE.