badrex/African-ASR

A repo for the code of training ASR transformer model for Kinyarwanda # African-ASR Develop Automatic Speech Recognition (ASR) Systems for African languages by fine-tuning multilingual speech models using CTC loss . This repository provides training scripts and configurations to build ASR models using 🤗 Hugging Face Transformers and related libraries. --- ## 🚀 Features - Training pipeline for ASR models - Configurable YAML files for flexible experiments - Language support for Kinyarwanda (initially) and extendable to other African languages - Support for wav2vec-BERT-2.0 and other multilingual models (XLSR, MMS, etc.) --- ## 📦 Requirements - Python 3.8+ - PyTorch (with GPU support recommended) - Hugging Face Transformers, Datasets, and Tokenizers - Other dependencies listed in `requirements.txt` Install the dependencies: ```bash pip install -r requirements.txt ``` ## ⚙️ Usage 1. Set up Hugging Face cache (optional): If you want to store downloaded models/datasets in a custom directory ```bash export HF_HOME="/path/to/huggingface/cache" ``` 2. Train the model Run the training script with a configuration file: ```bash python3 kinyarwanda-ASR/scripts/train_model.py \ --config kinyarwanda-ASR/config_files/ASR_train_config_sample.yaml ``` 3. In the configuration file, you should specifiy the base model (e.g., w2v-BERT-2.0), directory where the model will be saved, the training and validation datasets, as well as other hyperparameters such as the learning rate ```yaml # Project settings project: "Swahili-ASR" output_dir: "inprogress/swahili-ASR" seed: 4252 # Model settings pretrained_model: "facebook/mms-300m" #"ajesujoba/AfriHuBERT" #"facebook/w2v-bert-2.0" #"facebook/mms-300m" #"facebook/w2v-bert-2.0" freeze_feature_encoder: true # Training settings batch_size: 16 gradient_accumulation_steps: 2 num_epochs: 25 max_steps: 4000 learning_rate: 0.00007 warmup_ratio: 0.1 fp16: true gradient_checkpointing: true save_steps: 400 eval_steps: 400 logging_steps: 5 save_total_limit: 2 # Data settings # if use_custom_dataset is tru …