dsh-voice-input

Typing is slow, inconvenient, or simply unnecessary. For elderly users, newcomers to personal computing, or anyone who just wants to spare their wrists, voice input is not a productivity nicety — it is a usability prerequisite. Most off-the-shelf solutions mangle Chinese, demand a VPN, or ship your audio to a third-party server. dsh-voice-input solves all three inside the DSH chat window: tap the microphone, speak, and the text lands in the input box. No keyboard, no cloud, no external service.

Under the hood, faster-whisper runs as a persistent local process. The model loads once into memory; every subsequent transcription returns in milliseconds. Output is forced through OpenCC to simplified Chinese, so traditional characters never leak into the result. Anti-hallucination thresholds guarantee that silence stays silence — the UI reports no speech detected instead of Whisper inventing a paragraph from a cough. Models download from ModelScope domestic nodes with a direct connection, no VPN required. The default small tier stays around 500 MB. Want cloud speed? Drop a Groq or SiliconFlow API key into the config and the backend switches with zero code changes.

Who is this for? Parents fumbling through pinyin stroke by stroke, new computer users who freeze at the keyboard, remote workers with typing fatigue, and anyone who treats voice data as sensitive by default. The entire interaction is two taps — start, stop — with no modal dialogs and no setup wizard. A small white card appears under the microphone button, never blocking the chat. Tap, speak, done.