# 1. Clone the original fine-tuning repo
git clone https://github.com/saravananravi08/indicf5-finetune
cd indicf5-finetune

# 2. Install dependencies
uv sync

# 3. Download this model from HuggingFace
huggingface-cli download Saravananravi/indicf5-hinglish model.safetensors   --local-dir ./checkpoints

# 4. Generate speech
uv run evaluate --checkpoint-dir ./checkpoints

Usage Example

python

from indicf5_finetune.evaluate import load_model_from_checkpoint

model, step = load_model_from_checkpoint(
    checkpoint_dir="./checkpoints",
    vocab_path="path/to/vocab.txt"
)

# Synthesize
audio = model.generate("मैं आज office जा रहा हूँ")

Requirements

Python 3.10+
PyTorch 2.2+
f5_tts (bundled in the repo)
Vocos vocoder
CUDA GPU (recommended, works on CPU with reduced speed)

Model Details

Property	Value
Base Model	ai4bharat/IndicF5
Training	OpenSLR-104 (Hindi)
Steps	32,990
Language	Hindi + English (Hinglish)
Sample Rate	24kHz

License

Apache 2.0

Citation

bibtex

@software{indicf5_hinglish,
  title={IndicF5-Hinglish},
  author={Saravanan Ravi},
  year={2025},
  url={https://huggingface.co/Saravananravi/indicf5-hinglish}
}