SinhalaVITS-TTS-F2

SinhalaVITS-TTS-F2 - Female Voice 02

This is a specially trained Coqui TTS Coqui TTS model specially for Sinhala, developed by Dialog Axiata PLC and the Dialog – UoM Research Lab.

We trained it on a custom recorded dataset adapting a clear female voice.

---

Features

• Model architecture: VITS
• Language: Sinhala (si-lk)
• Training Sampling rate: 22050 Hz
• Framework: Coqui TTS

---

Dataset

• Voice: Female (Sanuki)
• Recording Sampling Rate: 44100Hz
• No. of Clips: 1096
• Total Length: >100mins (~2 hrs.)

Training Specs

• Hardware: NVidia GeForce GTX1060 6GB GPU
• Training Time: ~95 hours
• Global Steps: 190,000
• Batch Size: 16
• Epochs:
• Loss Convergence: Stable mel + KL losses

Installation

You can run this model locally using the included Flask-based inference server. This server will automatically use CUDA if it's available on your system.

1. First install requirements.

    pip install -r requirements.txt
   

2. Then start the API server

  python inference_F2.py

This starts a Flask server at http://localhost:8000.

1. Then you can use curl or any HTTP client (like Postman) to send Sinhala text to the server. The API endpoint is '/tts'

  curl -X POST http://localhost:8000/tts \
     -H "Content-Type: application/json" \
     -d '{"text": "ආයුබෝවන්"}' \
     --output output.wav

1. This API will,
   • Convert Sinhala text → Romanized Sinhala (via romanizer.py)
   • Generate speech using the VITS model
   • Return output.wav (Sinhala voice)

File Structure

  SinhalaVITS-TTS-M2/
    ├── Sanuki_190000.pth           # Fine-tuned VITS checkpoint
    ├── Sanuki_config.json          # Model configuration
    ├── romanizer.py                # Sinhala → Roman converter
    ├── inference_F2.py             # Flask-based inference server
    ├── requirements.txt            # Required dependencies
    ├── LICENSE                     # MPL-2.0 license
    └── README.md                   # This file

Contributors

• Kasun Ranasinghe (Dialog-UoM Reasearch Lab)
• Randika Silva (Dialog Axiata PLC)
• Vipula Wakkumbura (Dialog-UoM Reasearch Lab)

Acknowledgements

• PathNirvana (https://github.com/pathnirvana/coqui-tts) – Previous work in Sinhala TTS
• Coqui TTS – Open-source TTS framework enabling the foundation of this work
• Sinhala dataset contributor (Sanuki Bentharage) – for providing professional, quality speech samples

License

This model is released under the MPL-2.0 license.