---

✨ Why SpeakType

Most AI dictation tools are closed source, or route your voice through the vendor's own servers. SpeakType flips that:

🎬 Core experience

🎛️ Recognition engines (pick one, switch anytime)

1. Built-in offline recognition (default, recommended) — one-click model download inside the app: SenseVoice-Small for Chinese/English/Japanese/Korean/Cantonese (~0.27s per utterance, punctuation included), NVIDIA Parakeet TDT 0.6B v3 for the highest English + 25 European language accuracy, or whisper.cpp (tiny/base/small). No network, no account, no keys.
2. Any OpenAI-compatible /audio/transcriptions API — Base URL + API key + model name. Presets for OpenAI Whisper, Groq (free tier), Fireworks, Mistral Voxtral, SiliconFlow and Alibaba Bailian, with a connection test.
3. No-API-key web providers — ChatGPT web transcription (a free OpenAI account works) or Doubao voice, both reusing a session you sign into yourself inside the app. These use undocumented endpoints, are off by default, and may break or conflict with those services' terms — the account risk is yours to judge. See DISCLAIMER.md.

AI polishing likewise accepts any OpenAI-compatible chat endpoint (OpenAI, Google Gemini's OpenAI-compatible endpoint, Groq, DeepSeek, Zhipu GLM-4-Flash, Kimi, Qwen, or a local Ollama / LM Studio endpoint — for local endpoints the API key can be left empty). Without one, a local cleanup pass still handles self-corrections ("5pm — no, 6pm" → "6pm").

📦 Download & install

Latest release: https://github.com/wookat/speaktype/releases/latest · Website: https://speaktype.zalize.com

Or install via Scoop:

scoop bucket add speaktype https://github.com/wookat/scoop-speaktype
scoop install speaktype

1. Install (if SmartScreen objects, click "More info → Run anyway"; the installer is not commercially signed).
2. Settings → Speech → Built-in offline → download a model (or fill in your own API key).
3. Put the cursor in any input field, hold RightCtrl, speak, release.

🔒 Privacy boundary

• SpeakType has no servers — it collects and uploads nothing.
• Audio goes only to the service you configured; in offline mode it never leaves your machine.
• The "learn from my edits" comparison happens entirely on your machine.
• API keys, history and failed recordings live only in %APPDATA%\SpeakType.
• No third-party credentials are bundled in this repository.

🌏 Internationalization

Simplified Chinese, Traditional Chinese, English, Japanese and Korean built in (follows system or manual, applies instantly); the locale architecture welcomes more community additions.

🛠️ Contributing

cd desktop
npm install
npm run dev        # dev mode
npm run typecheck
npm run pack       # NSIS installer → release/

Stack: Electron + React 19 + Tailwind 4 + lucide-react; global hotkeys via uiohook-napi; typing via koffi SendInput; offline recognition via SenseVoice / Parakeet (sherpa-onnx) / whisper.cpp; enhanced VAD via Silero + onnxruntime; phone microphone via a Cloudflare Worker relay you can self-host. See desktop/README.md.

An earlier Chrome extension form lives in this repo too.

Issues and pull requests welcome — see CONTRIBUTING.md.

📄 License

MIT © wookat & SpeakType contributors

SpeakType is an independent open-source project, unaffiliated with OpenAI, Google, ByteDance, Zhipu or any other vendor. The no-API-key providers reuse a login session you create yourself against undocumented endpoints, which may not comply with those services' terms; the account risk is yours to judge. Prefer the built-in offline engine or your own API key if unsure. See DISCLAIMER.md.