A sequence-to-sequence model for Bangla speech recognition, implemented in PyTorch. The repository covers the full path from raw audio to a trained model: data loading, preprocessing, training and validation.
What's in it
- Seq2Seq model. An encoder–decoder architecture that transcribes Bangla speech to text.
- PyTorch throughout. The model, training loop and evaluation are all plain PyTorch.
- Data tooling. Utilities for loading, preprocessing and batching the speech dataset.
- Training & validation scripts ready to run against your own data.
Stack
- Python 3.7
- PyTorch
- Librosa — audio loading and feature extraction
- NumPy
Context
Bangla is spoken by hundreds of millions of people yet remains badly under-served by speech technology. This project grew out of that gap — and out of my undergraduate thesis, BanglaKotha: Bangla Automatic Speech Recognition Leveraging RNN-T.