All projects

Bengali long-form ASR

A documented progression from Whisper baselines to LoRA fine-tuning, VAD, chunked decoding, and multi-GPU inference for long Bangla recordings.

Status
Experiment archive
Role
Developed and documented the experimental pipeline with team Backprop Bandits.
Team context
Competition team project.
Speech research · Competition2026
  1. 01Long audio
  2. 02VAD + chunks
  3. 03Whisper LoRA
  4. 04Transcript
High-level flow only. Open the case study for decisions, evidence, and limitations.

Long, noisy Bengali recordings trigger silence, drift, repetition, and memory problems in straightforward Whisper inference.

  • 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.

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.

Next projectGPUShare