@wayneyu430227/dsh-voice-web

English | 中文

Dedicated /voice WebSocket carrier. Binary client frames are 16 kHz PCM microphone chunks; binary server frames are 24 kHz PCM speech. JSON frames carry readiness, identified ASR and output text, task observations, speech lifecycle, audio commit, response-specific interruption, playback completion, and explicit session.close. An unexpected socket loss detaches the browser so reconnecting the same durable Voice Session within reconnectGraceMs reuses its exact provider conversation; explicit stop or a malformed or oversized frame final-closes it. Provider termination closes the browser transport instead of leaving the UI attached to a dead conversation. The route reuses the browser Host/Origin trust fence and does not share the ordinary RPC downlink.

Model Experience

Browser voice transport

What the model sees

The package contributes zero model-visible text; /voice transports audio and provider-independent lifecycle events only.

Token effect

Zero tokens are added directly.

KV Cache effect

WebSocket frames do not modify a model request or its reusable prefix.

Known Limitations and Deferred Work

• Authentication remains the deployment's responsibility; trustedHosts is a DNS-rebinding and same-origin fence, not authentication.