Bengali-ASR

Bengali-ASR provides scripts and notebooks for building an automatic speech recognition (ASR) system for the Bengali language. The project uses the Whisper architecture with JAX/TPU as well as PyTorch implementations based on Wav2Vec2.

Project Structure

• setup.sh – installs Python 3.11 and all required packages (JAX, Transformers, PyTorch, etc.).
• download.sh – downloads training data from Kaggle and other public sources.
• functions.py, functions_infer.py – dataset utilities and dataloaders.
• run_train.py – main JAX training script for the Whisper model.
• run_train_txt.py – optional text-only training of the decoder.
• model_wav2vec_CTC.py – PyTorch approach using Wav2Vec2 with a CTC head.
• *.ipynb – Jupyter notebooks with experiments and evaluations.

Setup

1. Install dependencies

   bash setup.sh
   

2. Download data

   bash download.sh
   

   Place your kaggle.json credentials in the project root before running the script.

3. Activate the virtual environment

   source ~/.venv311/bin/activate
   

Training

Whisper (JAX/TPU)

Edit hyperparameters in run_train.py as necessary, then execute:

python run_train.py

Text-only fine-tuning

For additional training using only text data, run:

python run_train_txt.py

Wav2Vec2 CTC (PyTorch)

The file model_wav2vec_CTC.py contains a PyTorch implementation with a CTC loss. Run it directly after adjusting paths:

python model_wav2vec_CTC.py

Inference

functions_infer.py shows how to create an inference dataset and collate function. See the notebooks for end-to-end examples.

Notebooks

Open the Jupyter notebooks for exploration and evaluation:

jupyter lab

License

This repository is provided for research purposes. Please review the licenses of the datasets used before redistribution.