Switch from insanely-fast-whisper to WhisperX for transcription services,
gaining word-level granularity not available with FlashAttention in previous
API. Whisper was packaged in FastAPI image with basic diarization and
speaker-segment association capabilities.
Implementation prioritized speed to market over testing. Future iterations
will enhance diarization quality and add proper test coverage. API consumers
should expect behavioral differences and new capabilities.
The new endpoint requires title and content, breaking
the original implementation. This hotfix ensures staging works
immediately while I plan an LLM-based solution
for title generation.
Got an unexpected error while instantiating the openai client
since I installed sentry deps. I've upgraded openai version
without digging much, and it worked again.
This is written in a rush, boostrap the real logic with celery
worker to process meeting recording.
Lack of unit tests is critical, I am not proud of a it. I am
totally in a hurry for the demo. Not sure we will actually keep
this microservice.
lebaudantoine