Every ASR/TTS worker (faster-whisper, qwen3, vibevoice, xtts) kept every
model it ever loaded resident in GPU memory forever, and switching to a
different model (e.g. a different whisper size) just added another one
alongside it rather than freeing the old one. Combined with the shared
24GB GPU, this made memory pressure only ever go up.
Now: only one model stays resident per worker at a time (loading a
different model_id evicts the previous one first), and the whole cache
(plus, for faster-whisper, the diarization pipeline) is dropped after 2
minutes of no requests. Verified end-to-end: the idle timer actually fires
and reclaims memory, and switching qwen3 models evicts the old one.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Third ASR backend option alongside faster-whisper and qwen3. Builds
microsoft/VibeVoice from source (pinned to a specific commit, since it's
custom modeling code not in transformers' Auto* registry) into its own venv,
inheriting the base image's torch/CUDA. Comes with built-in speaker
diarization (VibeVoiceASRProcessor.post_process_transcription returns
per-segment speaker ids directly, no separate pyannote pass needed).
Verified end-to-end against a real recording: 200 OK, correct Korean
transcription, speaker labels populated.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
