dsh-voice-live

In the DSH interface, typing is the only input channel, but speaking is often faster and more natural in many workflows. dsh-voice-live closes that gap for DSH clients: tap the mic button beside the composer, speak, and the assistant replies aloud, all without switching input modes.

The core capability runs over a single WebSocket real-time channel: 16 kHz microphone PCM is fed to Volcengine streaming ASR (bigmodel_async with server-side VAD endpointing), recognized text streams incrementally into the composer draft, and assistant reply text is synthesized sentence by sentence via Volcengine seed-tts-2.0 TTS before playback. Barge-in lets you interrupt mid-reply, auto-reconnect with exponential backoff covers dropped connections (up to 5 attempts), 30 verified voices are available with live preview, and an optional reply-first mode has the assistant confirm verbally before queuing the full answer. The API key is read only on the host process; the browser never holds or transmits a key.

Best suited for zh-CN users who want hands-free DSH interaction: voice-first workflows, mobile typing-free input, or quick Q&A during meetings. Note that the plugin currently requires the DSH monorepo build chain for installation, and the wake-word feature relies on Web Speech API, which is only reachable via Safari on mainland networks.