MMS-300M Nepali Fine-tuning

This repository contains code for fine-tuning the MMS-300M model on Nepali speech data.

Setup

1. Install the required dependencies:

pip install -r requirements.txt

2. Prepare your dataset:

   • Place your Nepali audio files (WAV format) in a directory
   • Create a transcript file (JSON or CSV) with the following format:
     • For JSON: {"audio_file_id": "transcript text", ...}
     • For CSV: Two columns with headers "file_id" and "text"

3. Process your dataset:

python prepare_dataset.py

Edit the paths in the script to point to your audio files and transcript file.

Training

1. Start the training:

python train_mms_nepali.py

The training script will:

• Load the MMS-300M model
• Fine-tune it on your Nepali dataset
• Save checkpoints and the final model
• Log training progress to Weights & Biases

Model Output

The fine-tuned model will be saved in:

• Checkpoints: ./mms-nepali-model/
• Final model: ./mms-nepali-model-final/

Requirements

• Python 3.8+
• CUDA-capable GPU (recommended)
• At least 16GB RAM
• Sufficient disk space for the dataset and model

Notes

• The training script uses mixed precision training (fp16) for better performance
• Gradient accumulation is used to handle larger effective batch sizes
• The model is evaluated every 400 steps
• Training progress is logged to Weights & Biases for monitoring

Customization

You can modify the training parameters in train_mms_nepali.py:

• Batch size
• Learning rate
• Number of epochs
• Evaluation frequency
• etc.