dsh-koein

Voice input in DeepSeek Harness has usually meant a quick click on a mic button or a hotkey press before you could start talking. dsh-koein collapses that into a single spoken phrase: a resident keyword-spotting engine listens at roughly 50 times real-time speed, and the moment it hits, the pipeline switches to streaming recognition with live subtitles. No mouse, no keyboard, no extra plugin needed.

The recognition engine runs entirely on sherpa-onnx local inference. It handles Chinese and English together with automatic punctuation, at a 480 ms latency block that processes about 8 times real-time on a single CPU thread. Wake words can be any Chinese phrase, tokenized automatically against the model vocabulary with no retraining. The engines live in a forked child process, isolating onnxruntime DLL version conflicts and containing native crashes away from the host. Steady-state memory is around 330 MB. All audio stays on the machine; there is no cloud call, no API key, and no recording written to disk.

It is a good fit for prompt writers who compose long text in DSH and would rather speak than type, for privacy-first users who do not want speech sent to a service, and for developers working in offline or constrained environments. After a one-time download of about 159 MB of models, you simply say a phrase and start dictating.