Appolinairee/fongbe-asr

# Fongbe ASR Automatic Speech Recognition system for Fongbe (Benin) using parameter-efficient fine-tuning of Whisper with LoRA. ## Overview This repository contains code and documentation for fine-tuning OpenAI's Whisper model on Fongbe speech data using Low-Rank Adaptation (LoRA). The project investigates optimal parameter-efficient fine-tuning strategies for low-resource African languages. **Dataset:** 13,581 audio-transcription pairs (12.22 hours total) **Sources:** Laleye et al. (Zenodo) + pyFongbe corpus **Split:** 80% train / 10% validation / 10% test ## Quick Start ### Local Training ```bash git clone github.com cd fongbe-asr python -m venv venv && source venv/bin/activate pip install -r requirements.txt python scripts/download_sources_to_drive.py --dest data/raw python scripts/prepare_dataset_hf.py python scripts/finetune_whisper.py ``` ### Google Colab Use the provided `colab_training.ipynb` notebook for GPU-accelerated training with automatic GitHub/Drive synchronization. Use `scripts/download_sources_to_drive.py` to download the public source datasets directly into Google Drive or `data/raw`. See `WORKFLOW.md` for detailed setup instructions. ### Download Sources ```bash source .venv/bin/activate python scripts/download_sources_to_drive.py --dest /content/drive/MyDrive/fongbe/data/raw ``` ## Repository Structure ``` fongbe-asr/ ├── scripts/ │ ├── finetune_whisper.py # Main training script │ ├── experiment_lora_ranks.py # LoRA rank ablation study │ ├── download_sources_to_drive.py # Download public datasets to Drive/local │ └── prepare_dataset_hf.py # Dataset preprocessing ├── colab_training.ipynb # Colab notebook with sync ├── DATA_SOURCES.md # Verified dataset links and notes ├── TRAINING_GUIDE.md # Training documentation └── WORKFLOW.md # Development workflow guide ``` ## Research Questions 1. What is the op …