dsh-voice

An interview is a conversation, not a text thread. dsh-voice fills that gap for DeepSeek Harness (dsh) interview presets: the Agent speaks a question, the browser plays synthesized speech aloud, the candidate records an answer on the spot, and the transcript flows straight back into the conversation. The entire loop is turn-based by design—listen first, then speak—mirroring the natural rhythm of a real interview.

The plugin exposes three minimal tools: voice_speak, voice_listen, and voice_status, so the Agent invokes exactly what it needs. TTS and ASR engines are chosen independently: qwen and mimo on the cloud side, sherpa-onnx, piper, and whisper.cpp for fully offline local runs. An auto mode probes engines in priority order so a single outage never breaks the chain. Audio is written locally and served to the browser over a loopback route; cloud engines only ever receive the specific clip that needs transcription. API keys are injected via environment variables and never stored in config, keeping the privacy boundary tight.

Built for teams that fold dsh interview presets into their hiring workflow, and for independent developers who want voice I/O that can run entirely on-premises. If you are assembling an AI interviewer and need a stable, swappable, offline-capable voice channel, dsh-voice saves you from wiring up hundreds of lines of glue code.