VITS

Few TTS codebases have been forked as often as VITS, the ICML 2021 reference implementation of a conditional variational autoencoder with adversarial training and normalizing flows that made single-stage end-to-end speech synthesis practical. A stochastic duration predictor lets it speak the same text with varied rhythm and pitch, and at release it produced more natural audio than the two-stage text-to-spectrogram-plus-vocoder pipelines that were standard at the time. The repository provides training recipes for single-speaker LJ Speech and multi-speaker VCTK at 22050 Hz, pretrained checkpoints via Google Drive, and an interactive Colab inference notebook; setup requires PyTorch, espeak for phonemization, and compiling the Cython monotonic alignment search module. Training realistically requires a CUDA GPU, while inference of pretrained models can run on CPU. The MIT license allows commercial use. Its architecture underpins a large share of the modern open TTS ecosystem, including the Coqui, so-vits-svc, and Bert-VITS2 lineages, and the original repo remains a research baseline at roughly 7.9k GitHub stars.