dsh-voice-assistant

Typing in dsh web for long stretches means constant back-and-forth between keyboard and mouse, quietly eroding focus. dsh-voice-assistant places a microphone button and a sci-fi-style status bar right beside the input field: say the wake word (default: Xiao Jing) to activate, and your dictated words flow into the draft in real time while model replies are read aloud in a natural Chinese voice, all without touching a single key or clicking the mouse.

The engine defaults to sherpa-onnx WASM running entirely in the browser, giving you offline-capable speech recognition with zero dependency on Google services and no proxy needed on Chinese networks. Streaming intermediate results mean the wake word triggers the instant it leaves your mouth, and spoken commands like "execute send" or "execute clear" fire mid-utterance without waiting for a silence window. Two-layer pinyin tolerance (exact homophone hit plus near-homophone confirmation) guards against both missed wakes and false triggers, while the mandatory "execute" prefix ensures everyday words such as "send" or "delete" in your prose can never accidentally fire a command. Four layers of audio processing, including automatic gain control, high-pass filtering, tanh soft-clipping, and silent-block discarding, keep recognition stable across microphones and distances.

This plugin is built for developers and everyday users who type heavily in dsh web, care about audio privacy, or work in networks where Google services are restricted. Whether drafting long articles, issuing quick commands, or holding multi-turn conversations, saying one wake word and talking is an order of magnitude faster than typing.