MangoVoice
Push-to-talk dictation that actually speaks Traditional Chinese, not Simplified with a Taiwan label on it.
About
Hold a key, speak, release, and the text lands wherever your cursor already is. Whisper runs entirely on-device, so no audio leaves the machine, and the output is converted to Taiwan Traditional Chinese vocabulary instead of the Simplified text every general-purpose dictation tool produces.
Every dictation tool claims Chinese support. In practice that means a language code called zh, output in Simplified characters, mangled proper nouns, and mainland vocabulary in place of Taiwanese. MangoVoice solves exactly that one problem and nothing else. It runs Whisper large-v3-turbo locally, pipes every result through OpenCC's s2twp conversion so 软件 becomes 軟體 and 鼠标 becomes 滑鼠, and lets you feed your own jargon in as a prompt so Vercel stops being transcribed as Verso. Hold a key you pick, talk, let go, and the text is pasted into LINE, Notion, VS Code, a browser, anything with a cursor. It can also translate as you speak: a small language model runs locally next to Whisper, so you say one language and type another without an API call. No account, no API key, no subscription, and nothing on the network once the models are downloaded. Pay what you want, including nothing.
Highlights
- Hold one key to talk, release to type, into any app with a cursor
- Whisper large-v3-turbo runs on-device; audio never leaves the machine
- OpenCC s2twp conversion for genuine Taiwan vocabulary, not just Simplified-to-Traditional characters
- Custom prompt so your own proper nouns transcribe correctly
- Optional offline translation: speak one language, type another, still no network
- Single Rust codebase for macOS and Windows
