VITS TTS for Indian Languages

This repository contains a VITS-based Text-to-Speech (TTS) model fine-tuned for Indian languages. The model supports multiple Indian languages and a wide range of speaking styles and emotions, making it suitable for diverse use cases such as conversational AI, audiobooks, and more.

Model Overview

The model ai4bharat/vits_rasa_13 is based on the VITS architecture and supports the following features:

• Languages: Multiple Indian languages.
• Styles: Various speaking styles and emotions.
• Speaker IDs: Predefined speaker profiles for male and female voices.

Installation

pip install transformers torch

Usage

Here's a quick example to get started:

import soundfile as sf
from transformers import AutoModel, AutoTokenizer

model = AutoModel.from_pretrained("ai4bharat/vits_rasa_13", trust_remote_code=True).to("cuda")
tokenizer = AutoTokenizer.from_pretrained("ai4bharat/vits_rasa_13", trust_remote_code=True)

text = "ਕੀ ਮੈਂ ਇਸ ਹਫਤੇ ਦੇ ਅੰਤ ਵਿੱਚ ਰੁੱਝਿਆ ਹੋਇਆ ਹਾਂ?"  # Example text in Punjabi
speaker_id = 16  # PAN_M
style_id = 0  # ALEXA

inputs = tokenizer(text=text, return_tensors="pt").to("cuda")
outputs = model(inputs['input_ids'], speaker_id=speaker_id, emotion_id=style_id)
sf.write("audio.wav", outputs.waveform.squeeze(), model.config.sampling_rate)
print(outputs.waveform.shape)

Supported Languages

• Assamese
• Bengali
• Bodo
• Dogri
• Kannada
• Maithili
• Malayalam
• Marathi
• Nepali
• Punjabi
• Sanskrit
• Tamil
• Telugu

Speaker-Style Identifier Overview

Citation

If you use this model in your research, please cite:

@article{ai4bharat_vits_rasa_13,
  title={VITS TTS for Indian Languages},
  author={Ashwin Sankar},
  year={2024},
  publisher={Hugging Face}
}