ALI-OUALA/darija-tts-data-cleaning

WhisperX-aligned Darija TTS release packaging, validation, and reporting workflow # Algerian Darija speech-data preparation WhisperX-based preparation workflow for the CAFE Algerian Arabic/French code-switching `small-clean` subset. One command creates training-ready STT data, a stricter TTS subset, review/rejection manifests, and QC reports. ## Release snapshot The preserved release export contains 940 validated segments: 16 kHz mono, 6.597 s mean duration, 0.00% clipping, 99.30% mean alignment coverage, and 96.60% of segments at coverage ≥ 0.95. ## One-command run Install the reproducible Windows environment with Python 3.11: ```powershell python -m pip install -r requirements.windows-py311.lock.txt ``` Run against the live Hugging Face dataset: ```powershell python scripts/prepare_dataset.py ``` Run against a local CAFE archive instead: ```powershell python scripts/prepare_dataset.py --source .\cafe-small-clean.zip --output-dir .\prepared ``` On Windows, `prepare_dataset.py` automatically relaunches with the repository `.venv311` interpreter when `python` resolves to Python 3.12/3.13. The explicit launcher is also available: ```powershell .\scripts\prepare_dataset.cmd --source .\cafe-small-clean.zip --output-dir .\prepared ``` The command processes only `small-clean`. It downloads Croissant metadata, reads the subset Parquet file directly, downloads only its audio records, then runs the existing WhisperX alignment and segmentation pipeline. Fresh WhisperX alignment is CPU/GPU work; use the Colab notebook when local resources are limited. Local Windows remains useful for smoke tests, validation, or repackaging a matching existing `tts_export`. ## Output ```text prepared/ ├── stt/accepted/metadata.csv # training input: audio|text ├── stt/review/ # held-out review samples ├── stt/rejected/ # rejected source/segment records ├── tts/accepted/metadata.csv # stricter TTS training input ├── tts/review/ ├── tts/rejected/ ├── source/ # normalized source audio + ma …