ethiomark/amharic-tts-poc

# Amharic TTS Prototype (WAXAL-based) ## Approach summary (before implementation) 1. **Inspect first, assume nothing**: detect real schema/config names and split names from `google/WaxalNLP` before preprocessing. 2. **Prototype-first pipeline**: keep an end-to-end path that runs on CPU with a toy backend, while allowing a stronger optional Coqui inference path. 3. **Defensive coding**: dynamically detect columns/splits and continue inspection even when some dataset configs fail. 4. **Clear fallback behavior**: API/UI/CLI should remain usable when checkpoints or heavy dependencies are missing. ## Key assumptions - WAXAL config names and schema may evolve; scripts are resilient but still depend on remote dataset availability. - High-quality Amharic TTS requires stronger modeling than the toy backend (e.g., Coqui or similar), more training time, and language front-end tuning. ## Project structure ``` . ├── README.md ├── requirements.txt ├── .gitignore ├── configs/ │ ├── train.yaml │ └── infer.yaml ├── scripts/ │ ├── inspect_waxal.py │ ├── preprocess.py │ └── train.py ├── app/ │ ├── api.py │ ├── audio_utils.py │ ├── dataset_utils.py │ ├── gradio_app.py │ ├── infer.py │ ├── model_loader.py │ └── text_normalization.py └── tests/ ``` ## Setup ```bash python -m venv .venv source .venv/bin/activate pip install -r requirements.txt ``` Optional stronger backend: ```bash pip install TTS ``` ## 1) Inspect WAXAL dataset schema ```bash python scripts/inspect_waxal.py ``` This script: - enumerates dataset configs, - discovers available split names per config, - auto-selects a text split (`train` preferred), - auto-detects text/audio columns, - prints 3 sample records, - writes `artifacts/dataset_summary.json`, including per-config errors. ## 2) Preprocess Amharic subset After inspection, choose the Amharic config: ```bash python scripts/preprocess.py \ --config \ --out-dir data/processed/amharic_tts \ --sample-rate 22050 \ --min-sec 0.5 \ --m …