dsh-guarded-live-voice

DSH previously offered only a simple push-to-talk transcription button with no bidirectional audio loop, no session-scoped consent binding, and no transport fences. DSH Live Voice fills that gap by building around a single bounded voice turn: you speak, you hear synthesized audio back, you review both transcripts, and you decide whether to place your own words into the current draft. It never submits the composer and never calls tools on your behalf, keeping full execution authority in your hands.

The experience is grounded in a guarded-consent security model. Consent is explicitly bound to the exact live DSH Session and Workspace; credential resolution happens only after a user-visible acceptance gesture; and the transport layer enforces loopback, same-origin, trusted-Host, and connection-count fences. Protocol parsing is fail-closed, and socket, timer, and session state are torn down deterministically at the end of every turn. On the input side, browser audio is downmixed and resampled to PCM16 mono at 16 kHz; on the output side, 24 kHz playback is ordered and rate-limited, with backpressure ceilings and timeout circuit-breakers on both paths to prevent resource leaks.

It is aimed at users and developers who want a bidirectional voice workflow inside DSH Web while requiring that every authorization step is visible and that the permission boundary is auditable. If you need continuous conversation, barge-in, or voice-triggered tool execution, this is not the goal. DSH Live Voice deliberately confines itself to a single-turn, draft-only interaction so that the security architecture lands before feature scope expands.