Problem
Long, noisy Bengali recordings trigger silence, drift, repetition, and memory problems in straightforward Whisper inference.
Contribution
- Combined Silero VAD, Whisper-sized windows, conservative decoding, and repetition cleanup.
- Tracked WER, CER, and real-time factor instead of training loss alone.
- Documented that the 500-step checkpoint is useful for pipeline validation, not a competitive final result.
Case note 01
The result that matters
The experiment archive preserves failures and fixes. Its current checkpoint reports roughly 88% evaluation WER, so it is presented as honest research iteration—not as a solved transcription system.