dsh-realtime-voice

Realtime voice in a DeepSeek Harness workflow has long been an either/or: either spin up Docker containers and local models, or point a browser at multiple voice APIs and lose the session scheduling, tool orchestration, and memory that Harness provides. dsh-realtime-voice collapses that pipeline into three steps—the browser captures PCM, the Harness Host proxies ASR and TTS over a same-origin WebSocket, and complete utterances flow through the native input box into the current session. No Python, no local weights, no extra agent service; restart Harness and the microphone is ready.

Three design principles anchor the core capability. First, every transcript and every spoken reply passes through Harness native input box and send flow, so reasoning, memory, web access, plugins, and tool dispatch all stay inside Harness; the voice vendor never gets a standalone turn. Second, a strict TurnCoordinator assigns a unique ID to each round, serialises ASR state events, and guarantees that a late callback from an older round can neither overwrite the new round nor trigger duplicate playback; new speech captured while TTS is playing is appended to the input box, and any keyboard edit, paste, clear, or connection drop cancels the auto-send to prevent mid-sentence commits. Third, an adaptive floor manager generates up to three natural bridging sentences when the wait exceeds 800 ms and cancels them the moment a fast answer arrives, keeping perceived latency smooth. The Qwen line supports Tencent voice-print soft gating and barge-in echo verification; the OpenAI line defaults to gpt-realtime-2.1; both share a single TTS writer queue.

It is aimed at developers and teams who want to say one sentence and drive a full Harness workflow—no local inference stack to babysit, no parallel voice-only session to manage. One install command plus the API key for the chosen line turns the microphone into a natural entry point for the entire workflow.