Changelog

User-visible changes only. Internal refactors and infrastructure work omitted.

v0.0.1 — Initial rollout#

The current phase ships the speech-to-text surface end-to-end. Module identity, permissions, admin pages, and the lifecycle hooks are all wired; the API mounts at /v1/modules/scaiecho/.

• Batch transcription via POST /transcribe. Inline for short audio, async via the arq worker pool for long audio with GET /transcribe/jobs/{id} for polling.
• Real-time WebSocket streaming at WS /stream/transcribe. Bearer auth from query or header, binary audio in, JSON delta frames out.
• WebRTC streaming under /stream/transcribe/webrtc/. Session create, SDP offer/answer, ICE trickle, control WebSocket for deltas, session teardown.
• Speaker library — list, get, enroll, warm-inspect, proactive warm, update, delete. Enrollment captures reference and consent audio with quality preflight and fan-out to every online pyannote node.
• Optional speaker diarization on streaming routes via diarize=true. Silently no-op on Backend B.
• Per-tenant backend policy at /tenant-policy. Lazy-provisioned from tier defaults.
• MCP tool scaiecho.transcribe registered in the catalog.
• Speaker erasure with immutable audit row and replica eviction.