dsh-voice-mode
A full-duplex voice plugin for DeepSeek Harness that streams speech recognition into an editable draft as you talk, reads replies aloud sentence by sentence with live captions, and lets you interrupt playback just by speaking.
dsh-voice-mode adds full-duplex voice interaction to a DeepSeek Harness session. Speaking produces streaming captions in near real time, with the message auto-sending after about 1.5 seconds of silence; the agent's reply is read aloud sentence by sentence with synchronized live captions, and speaking again while it plays back interrupts the reading immediately (true barge-in) so the user's words are heard right away.
Speech recognition runs on-device (streaming zipformer2 plus SenseVoice for the final pass) with no API key required, with inverse text normalization for numbers, dates, and currency, and an optional wake word with a whitelist for filler prefixes like "um" so they do not falsely trigger it. Text-to-speech defaults to Edge's cloud voices for speed and naturalness, with local VITS (Chinese) or Kokoro (bilingual) available for privacy, and a "yield" mode lets short backchannel words pass without triggering a hard interruption.
It is installed as a DeepSeek Harness plugin and includes 380 passing tests across 28 suites.