High-Logic/Genie-TTS

What it solves

GENIE is a lightweight inference engine designed to make high-quality speech synthesis (TTS) fast and efficient, specifically targeting CPU performance. It removes the heavy dependencies of the original GPT-SoVITS project, allowing for near-instantaneous audio generation without requiring a powerful GPU.

How it works

It builds upon the GPT-SoVITS V2 and V2ProPlus models, converting them into an optimized ONNX format. By doing so, it reduces the runtime size and first-inference latency compared to both the official PyTorch and ONNX implementations. It also provides a simplified Python API for loading character voices, setting reference audio for emotion cloning, and running synthesis.

Who it’s for

Developers and users who want to integrate high-quality, multi-language TTS into applications without the overhead of large ML frameworks or the need for dedicated GPU hardware.

Highlights

• CPU Optimization: Significantly lower first-inference latency and smaller runtime memory footprint compared to official models.
• Multilingual Support: Supports Japanese, English, Chinese, and Korean.
• ** uma API**: Simple pip install and a few lines of code to generate speech from predefined or custom characters.
• Model Conversion: Includes built-in tools to convert original .pth and .ckpt GPT-SoVITS models to the optimized ONNX format.
• API Server: Comes with a built-in FastAPI server for easy deployment as a web service.