A minimal implementation of the accent-origin classifier from the UCLA M214A Project to detect contemporary and historical accents from several US cities based on the CORAAL database (cf. https://oraal.github.io/coraal). This implementation uses only basic audio features and XGBoost. No embeddings, no neural networks, pure decision-tree extravaganza.
The training pipeline is available in the main.ipynb file. Intermediary files
and plots available in the subfolder.
The helper functions assume the following directory structure:
- project_data/
- train/
- sample1.wav/
- sample2.wav/
- test_clean/
- sample1.wav/
- sample2.wav/
- test_noisy/
- sample1.wav/
- sample2.wav/
- features/