This repository provides an unofficial PyTorch implementation of VITS2, a single-stage text-to-speech (TTS) model designed for improved naturalness, speech characteristic similarity, and computational efficiency. It targets researchers and developers seeking to build or fine-tune advanced TTS systems, offering a fully end-to-end approach that reduces reliance on external phoneme conversion.

How It Works

VITS2 enhances its predecessor by incorporating several architectural improvements. It features a transformer block within the normalizing flow for better sequence modeling, a speaker-conditioned text encoder for multi-speaker synthesis, and a duration predictor with adversarial loss and noise-scaled monotonic alignment search for more robust duration modeling. These components contribute to a more natural and efficient speech synthesis process.

Quick Start & Requirements

• Install: Clone the repository and install requirements from requirements.txt.
• Prerequisites: Python >= 3.10, PyTorch 1.13.1+, espeak.
• Data: Download LJSpeech or VCTK datasets and create symbolic links. Preprocessing scripts are provided.
• Training: Use train.py for single-speaker and train_ms.py for multi-speaker models, referencing configuration files.
• ONNX Export: Scripts export_onnx.py and infer_onnx.py are available.
• Links: Discussion Page for logs and community contributions.

Highlighted Details

• Implements key VITS2 features: transformer normalizing flow, speaker-conditioned text encoder, and adversarial duration predictor.
• Supports ONNX export for efficient inference.
• Includes Gradio demo support.
• Offers pretrained checkpoints for LJSpeech.

Maintenance & Community

The project is actively maintained by the author, with contributions and discussions welcomed via GitHub issues. Special mentions are given to contributors for feedback, guidance, and resource support.

Licensing & Compatibility

The repository does not explicitly state a license in the README. Users should verify licensing for commercial or closed-source use.

Limitations & Caveats

The implementation is unofficial and may not perfectly mirror the original VITS2 paper's exact configurations or performance. Some advanced features might still be under development or require expert verification.