Fish Diffusion

Fish Diffusion is a state-of-the-art, professional-grade neural audio synthesis framework built on PyTorch, specifically optimized for high-fidelity Singing Voice Conversion (SVC) and Singing Voice Synthesis (SVS). By 2026, it has emerged as the definitive successor to traditional Diff-SVC implementations, offering a modular architecture that integrates diffusion probabilistic models with advanced vocoders like HiFi-GAN and BigVGAN. The system excels in capturing nuanced vocal textures, vibrato, and emotional delivery that typical VITS-based models often flatten. Its technical core utilizes RMVPE (Robust MVP-based Pitch Estimation) for ultra-accurate F0 extraction, ensuring that pitch tracking remains stable even in complex polyphonic backgrounds. Positioned as the 'Stable Diffusion of Audio,' the framework allows researchers and studio engineers to train custom voice models with as little as 30 minutes of clean audio. It supports multi-speaker training, cross-lingual synthesis, and shallow diffusion techniques which significantly reduce inference latency without sacrificing the 44.1kHz studio-quality output. The project is maintained by Fish Audio, providing a robust bridge between open-source community innovation and commercial-grade reliability.