dsh-voice

Typing is not the only way to put words into a DeepSeek Harness session. dsh-voice places a small microphone button right before the send key: tap to record, tap again to stop, and the transcript lands directly in the input box. It turns speaking into a first-class input method, as natural as tapping send.

Three recognition engines are available and swappable with a single constant. The browser-native Web Speech API works out of the box with zero setup; iFlytek listening delivers the best Chinese accuracy with near-instant results while keeping credentials on your local server; and a local Whisper (small q8) model runs fully offline so audio never leaves the machine. On top of that, the plugin adds a microphone device picker, automatic gain for quiet mics, a local gate that discards recordings shorter than 0.25 seconds to save cloud quota, and ordered stitching of multi-segment iFlytek results so long sentences are never truncated.

If you use DeepSeek Harness to write long-form text, take notes, or think out loud and prefer speaking over typing, dsh-voice fits right in. Its liquid-glass styling uses the same design tokens as the dsh-liquid-glass theme, so it feels like part of the interface rather than an add-on.