Speech Recognition Model

This repository contains a speech recognition model implementation using PyTorch. The model uses CTC (Connectionist Temporal Classification) loss for training and beam search decoding for inference.

Project Structure

.
├── README.md
├── requirements.txt
├── model.py           # Model architecture
├── dataset.py         # Data loading and preprocessing
├── train.py          # Training script
├── utils.py          # Utility functions
└── train.ipynb       # Training notebook

Setup

1. Install dependencies:

pip install -r requirements.txt

2. Install ctcdecode:

git clone --recursive https://github.com/parlance/ctcdecode.git
cd ctcdecode
pip install .
cd ..

Training

You can train the model either using the Python script or the Jupyter notebook:

Using Python script:

python train.py

Using Jupyter notebook:

1. Open train.ipynb
2. Adjust the configuration parameters if needed
3. Run all cells

Model Architecture

The model uses a bidirectional LSTM architecture with the following components:

• Multiple LSTM layers
• Dropout for regularization
• Linear projection layer
• CTC loss for training
• Beam search decoding for inference

Configuration

The main configuration parameters are:

• input_dim: Input feature dimension
• hidden_dim: LSTM hidden dimension
• num_layers: Number of LSTM layers
• num_classes: Number of output classes
• dropout: Dropout rate
• batch_size: Training batch size
• learning_rate: Initial learning rate
• epochs: Number of training epochs
• patience: Patience for learning rate scheduler
• beam_width: Beam width for decoding

Data Format

The model expects:

• Features in .npy format
• CSV files with training/validation data information
• Directory structure:

  data/
  ├── features/
  │   ├── sample1.npy
  │   ├── sample2.npy
  │   └── ...
  ├── train.csv
  └── val.csv