This project provides state-of-the-art Text-to-Speech (TTS) models for 13 Indian languages, addressing the under-representation of these languages in speech synthesis research. It targets developers and researchers working with Indian languages, offering improved speech quality over existing models.

How It Works

The system utilizes a unified architecture based on FastPitch for acoustic modeling and HiFi-GAN V1 for vocoding. Models are trained jointly on male and female speakers, a configuration identified through extensive evaluation of acoustic models, vocoders, loss functions, and training schedules. This approach yields superior performance across Dravidian and Indo-Aryan languages.

Quick Start & Requirements

• Install: Clone repositories for Trainer and TTS, install dependencies using pip3 install -e .[all] for both, and then pip3 install -r requirements.txt.
• Prerequisites: CUDA 11.3, PyTorch, libsndfile1-dev, ffmpeg, enchant.
• Resources: Requires significant computational resources for training. Inference setup involves downloading pre-trained model weights and configuration files.
• Links: ArXiv Preprint, Audio Samples, Try It Live, Video.

Highlighted Details

• Supports 13 Indian languages: Assamese, Bengali, Bodo, Gujarati, Hindi, Kannada, Malayalam, Manipuri, Marathi, Odia, Rajasthani, Tamil, and Telugu.
• Achieves significant improvements over existing models as measured by Mean Opinion Scores (MOS).
• Models are open-sourced on the Bhashini platform.
• Code references the coqui-ai/TTS library.

Maintenance & Community

• Developed by AI4Bharat, a mission-driven initiative.
• Accepted at ICASSP 2023.

Licensing & Compatibility

• The README does not explicitly state a license. The underlying coqui-ai/TTS library is typically Apache 2.0, but this specific fork's license requires verification.

Limitations & Caveats

• The setup instructions involve manual patching of the Trainer library, indicating potential instability or ongoing development.
• The license is not clearly specified, which may impact commercial use.