Bengali Long-Form Speech Recognition (DL Sprint 4.0)

This project is for the Kaggle competition: DL Sprint 4.0: Bengali Long-Form Speech Recognition.

Our solution pipeline:

1. Used arijitx/wav2vec2-large-xlsr-bengali to create timewise-annotation chunks from long Bengali audio so they fit Whisper training/inference windows.
2. LoRA fine-tuned the pretrained bengaliAI/tugstugi_bengaliai-asr_whisper-medium model on those prepared chunks.
3. Ran final inference with the tuned model for long-form Bengali ASR.

Main files:

• finetuning-whisper-0-37.ipynb: model fine-tuning workflow
• six-seven-inference-inferasr.ipynb: inference and decoding pipeline
• wav2vec2json.py: helper script for JSON-style output preparation