README_SETUP.md

Local setup (HuggingFace Spaces friendly)

This project expects heavy model checkpoints to live under checkpoints/ (not bundled in the repo).

Install Python dependencies (recommended in a virtualenv):

python -m pip install -r requirements.txt

Run the UI locally for debugging:

python webui.py --verbose

Notes for Spaces deployment:

• Do not add model checkpoints into the repository. Place them in checkpoints/ and configure config.yaml as needed.
• The app sets HF_HOME, TRANSFORMERS_CACHE and TORCH_HOME to checkpoints/hf_cache by default to keep downloads within the workspace.

Hindi support:

• The text normalization layer contains an Indic phoneme pipeline preserving Devanagari akshara (matras, halant, nukta, anusvara). If indic_nlp is available it will be used to tokenize words; otherwise a safe fallback segmentation is applied.