dsh-live-voice

DSH Live Voice solves one specific problem: you want to speak a single voice turn inside a DSH session, hear the audio response from the AI, review both transcripts, and optionally drop your text into the current draft — without the tool ever making decisions for you. It is not a speech-to-text microphone button. It is a bounded, single-turn voice interaction: bidirectional native audio, session-bound consent, and an explicit second confirmation before any transcript enters the draft.

The core design revolves around control. On the input side, one PCM16 mono 16 kHz recording turn capped at 30 seconds; on the output side, PCM16 mono 24 kHz playback capped at 60 seconds. The Qwen audio handshake requires explicit pcm codec, text-plus-audio modalities, and manual turn detection — any tool invocation or non-message capability is rejected outright. Authority is revalidated before every audio append and commit; session-object reuse or workspace moves fail closed immediately.

Who it is for: developers already running a DSH Web profile who want to experiment with voice interaction in a guarded context. Preview.5 ships with bundled synthetic audio and fixed transcripts, so you can complete a full 24-second walkthrough in two minutes without Qwen credentials, a physical microphone, or an external server. It does not promise continuous conversation, barge-in, or composer submission — it is exactly one turn, done.