dsh-realtime-voice

If you spend your days talking to a DSH Agent, you have probably hit this wall: you want to add a step, tweak a parameter, or just ask how things are going, but you have to pull your headphones off, hover your fingers over the keyboard, and type a whole sentence before hitting Enter. dsh-realtime-voice exists to remove that friction. It injects a Qwen Audio Realtime voice channel directly into the built-in DSH 3080 WebUI, placing a dial button right next to the send key. No new page, no extra port, no source changes. Pick up, talk, interrupt when you need to, hang up when done.

Under the hood it is a dual-plane architecture. Qwen Audio Realtime handles all listening, speaking, and natural conversation at millisecond latency; the moment a request touches real work such as file operations, application calls, web search, or printing, four narrow semantic Function Calls (handoff, cancel, approval, question) route the full intent to the Agent model bound in the current DSH session. The plugin also subscribes to DSH progress events, approval prompts, follow-up questions, and terminal states, feeding them back into the voice channel for spoken confirmation. Together with a draggable voice float, full-duplex real-time interruption, a three-layer acoustic barge-in strategy, a thirty-second reconnection lease, and a shared protocol used by both the WebUI and a WeChat mini-program, a long call stays intact even when the network stutters.

It is built for developers and technical users who already run long tasks, multi-tool pipelines, or multi-step Agent workflows inside DSH. No redesign of your interaction flow is needed. Put your headphones on, replace typing with talking, and let voice become the most natural line between you and your Agent.