oza75/bambara-whisper-asr-finetuning

# Bambara Whisper ASR Fine-Tuning This repository contains scripts for fine-tuning the OpenAI Whisper model for automatic speech recognition (ASR) in Bambara, though the approach can be adapted to other languages. ## Overview The provided scripts enable users to fine-tune Whisper models on a specific language dataset, focusing on improving ASR capabilities. This implementation leverages the Hugging Face `transformers` library and is designed to be flexible, allowing modifications to fine-tune on other languages and datasets. ## Features - CLI-based configuration for easy adjustment of training parameters. - Integrated logging for both console and file output to track the training process. - Modular design for easy customization and extension. ## Prerequisites Before you begin, ensure you have the following installed: - Python 3.10 or later ## Installation First, clone this repository to your local machine: ```bash git clone github.com cd bambara-whisper-asr-finetuning ```` Then, install the required Python packages: ```bash pip install --upgrade -r requirements.txt ``` ## Usage To start fine-tuning the Whisper model, use the `main.py` script. The script supports various command-line options to customize the training parameters. ### Basic Usage ``` python main.py ``` ### Advanced usage You can specify training parameters using command-line arguments. Here are some examples: ``` accelerate launch --config-file accelerate_config.yaml main.py --deepspeed_config deepspeed_config.json --fp16 --model_checkpoint openai/whisper-medium --learning_rate 6.25e-06 --per_device_train_batch_size 64 --gradient_accumulation_steps 1 --per_device_eval_batch_size 32 --output_dir whisper-bambara-asr-002 --push_to_hub ``` ### Available Command-Line Options Available options for detailed tuning and configuration are listed in the config.py. ## Configuration Modify the config.py to change default settings or add new p …