- Szycy
- 24 września 2026 03:50 UTC
- Autor
- Kamo
- Pochęt się
- 354e7eb
/transcribe read the upload with getBytes() on a 2Gi pod, and a meeting recording runs past a gigabyte. It now copies the spooled upload to the scratch volume and ffmpeg reads that file directly (transcodeToWav16kMono(Path)); only the WAV comes back as bytes. The copy is deleted whether transcription succeeds or not. The whisper call's read timeout was ten minutes, shorter than CPU transcription of a long meeting takes. It now matches transcribeToVtt's two hours, and posts the WAV unbuffered so it is not copied into heap a second time.