Ecosystem overview for everything related to generative-models.
VoiceFlow is an efficient Text-to-Speech (TTS) system based on Rectified Flow Matching, addressing the efficiency limitations of traditional diffusion models in speech synthesis. As the official implementation of its ICASSP 2024 paper, it generates high-quality mel-spectrograms by learning a continuous flow between noise and data. Through a flow rectification process, it further optimizes the sampling trajectory, achieving superior synthesis quality and efficiency with a limited number of sampling steps. It features Kaldi-style data organization, flexible training configurations, supervised duration modeling, and experimental voice conversion capabilities.