dsh-voice: A DSH plugin that brings two-way voice conversation to DeepSeek Harness

What Does It Do?

It is a two-way voice DSH plugin for DeepSeek Harness that turns the DSH web app into a hands-free conversation loop: you speak, the agent replies out loud, and the microphone keeps listening. It solves the problem of typing every turn and reading every reply, folding voice input, spoken output and auto-read into a single plugin. Maintained by martindye under the MIT license and written in TypeScript, it works on any device the web app runs on, including iPhone and PC.

Core Features

• Unified composer button: voice entry on an empty draft, a send arrow while typing, a stop square while the agent runs or the system speaks, and a cross during an active voice turn.
• Continuous listening: the microphone reopens on its own after a reply ends, and an auto-open that captures no words is quietly re-offered a few times.
• Spoken pause commands: saying "hold the mic", "pause the mic" or "stop listening" parks the listening loop without sending a message.
• In-call mute: during a voice turn the dictate button becomes a mute/unmute toggle, and muted audio never reaches the recognizer.
• Per-message and automatic read-aloud: every finalized assistant message gets a speak/stop button, and fresh replies younger than two minutes are read aloud on a device that has auto-read enabled and has used the voice composer.
• Per-device auto-read toggle: a control in the session header corner, persisted per browser, with a green dot when the device is voice-armed.

How to Use This Plugin?

After enabling it in DSH, open the web app and use the unified voice button beside the composer: tap it on an empty draft to start speaking, and the turn sends when you finish. Tap the speaker button on any assistant message to hear it read aloud. To have new replies spoken automatically, turn on the auto-read toggle in the session header corner; the setting applies only to that browser. Speech is synthesized by the plugin's own TTS service, spawned on demand on the machine running DSH, with text resolved from the session log and stripped of code, links and markdown first; the engine can be the GPU qwen3 or the CPU pocket fallback, defaulting to qwen3 when its models are present.

How to Troubleshoot This Plugin?

If speech synthesis does not work, first confirm that Python 3 is on the PATH of the machine running DSH and that numpy and requests are installed; the CPU fallback engine needs its additional speech dependency as well. If the GPU engine is unavailable, check that the models sit in the expected model directory and that the engine selection file points at the right engine. When the port is taken or the service fails to start, inspect the host log for the voice service and adjust the port or restart the service as described in the official instructions.