dsh-voice

Writing long prompts by hand in the DeepSeek Harness Web UI is slow, tiring, and mangles fleeting ideas. dsh-voice drops a full voice-input pipeline straight into the chat surface: streaming dictation slices speech at natural breath pauses and types live into the composer, voice messages ship with a cancel countdown, and push-to-talk works on both mouse (hold the wave button) and keyboard (hold Ctrl), so your hands stay off the layout entirely.

What sets it apart is the multi-engine fallback chain. The browser Web Speech API delivers zero-latency local recognition with floating captions; Deepgram, Groq, HuggingFace, OpenAI, SiliconFlow, and other cloud backends hot-swap in and out of the path; whisper.cpp and SenseVoice-ONNX / Sherpa-ONNX guarantee fully offline, private transcription with no network dependency. The moment any tier runs dry or hits a 429, requests glide to the next one and the audio stream never drops. Code variables and identifiers pulled from your draft are injected into the STT context to sharpen recognition of technical jargon, and the transcribe_audio agent tool lets models ingest recordings straight from disk. A WebSocket realtime bridge also supports OpenAI Realtime API for low-latency bidirectional audio.

It is built for three profiles: heavy Harness users who compose long prompts and crave voice over typing; developers and teams with strict data-privacy requirements who need fully offline or intranet voice recognition; and production environments that demand redundant cloud-plus-local STT coverage so voice input never goes dark.