LunaVox is a lightweight inference engine designed for near-instantaneous speech synthesis, built upon the GPT-SoVITS V2 architecture. It targets users seeking high performance and convenience in TTS, offering significant improvements in speed and resource utilization, especially on CPU hardware. The project provides an optimized, self-contained solution for deploying TTS models efficiently, making advanced speech synthesis more accessible.

How It Works

LunaVox functions as a streamlined inference engine, integrating TTS inference, ONNX model conversion, and an API server into a cohesive package. Its core innovation lies in optimizing the original GPT-SoVITS model for exceptional CPU performance. This optimization results in substantially lower latency and drastically reduced model and runtime sizes compared to official PyTorch and ONNX implementations. This approach prioritizes out-of-the-box usability and efficiency, making it suitable for deployment on a wider range of hardware.

Quick Start & Requirements

Installation is straightforward via pip: pip install lunavox-tts. Python version 3.9 or higher is required. Windows users may encounter installation challenges with the pyopenjtalk dependency, which necessitates installing Visual Studio Build Tools with the "Desktop development with C++" workload. A convenient quick_tryout_en.py script is provided for immediate testing without local model files, automatically downloading necessary components.

Highlighted Details

• Performance: Achieves a first inference latency of 1.13s on an i7-13620H CPU, outperforming official PyTorch (1.35s) and ONNX