vitthal-bhandari/low-resource-asr

Mozilla Common Voice Spontaneous Speech ASR for Underrepresented Languages # Low-Resource ASR ASR for Endangered Languages using Mozilla Common Voice Spontaneous Speech datasets. ## Overview This project evaluates and improves automatic speech recognition (ASR) systems for 21 underrepresented languages from Africa, Asia, Europe, and the Americas. It establishes baselines using MMS and Whisper, implements improved architectures, and conducts linguistic error analysis. ## Setup ### Prerequisites - Python 3.10+ - uv package manager ### Installation ```bash # Install uv (if not already installed) curl -LsSf astral.sh | sh # Clone the repository git clone github.com cd low-resource-asr # Create virtual environment and install dependencies uv sync # For development dependencies uv sync --all-extras ``` ### Activate Environment ```bash # Activate the virtual environment source .venv/bin/activate ``` ## Project Structure ``` low-resource-asr/ ├── data/ │ ├── mozilla_speech_data/ # Mozilla Common Voice datasets │ └── linguistic_resources/ # Grammars, phoneme inventories ├── src/ │ ├── data/ # Data loading and preprocessing │ ├── models/ # Model definitions and configs │ ├── training/ # Training scripts │ └── evaluation/ # Evaluation and error analysis ├── notebooks/ # Jupyter notebooks for exploration ├── scripts/ # CLI scripts ├── results/ # Experiment results ├── .notes/ # Project documentation ├── pyproject.toml # Dependencies (uv) └── README.md ``` ## Languages 21 languages across 4 regions: | Region | Languages | |--------|-----------| | Africa | Bukusu, Chiga, Nubi, Konzo, Lendu, Kenyi, Thur, Ruuli, Amba, Rutoro, Kuku | | Americas | Wixárika, Southwestern Tlaxiaco Mixtec, Michoacán Mazahua, Papantla Totonac, Toba Qom | | Europe | Gheg Albanian, Cypriot Greek, Scots | | Asia | Betawi, Western Penan | …