MedComms: VI ⇄ EN speech translation pipeline Diagram

Two independent directions

vi-en and en-vi each get a Pipeline: 3 spawned processes linked by queues
Each direction serves one conversation; a second session is closed with code 1013

Partial and final

The segmenter emits growing segments, so the translation improves while someone speaks
One final per finished sentence; the front door numbers it (id) for ratings and corrections

Waking the engine

Sign-in makes the front door ping the engine; GET /api/engine reports ready / starting / unavailable
Modal runs at most 1 container, idle after idle_minutes; models cached in the medcomms-cache volume

Graceful stop and control

Stop sends {"type": "end"}: the engine finishes the sentence in progress (final + END)
medcomms cloud start|stop keeps model settings in a Modal Dict, then runs modal deploy