VITS is the official implementation of the paper “Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech”, one of the most influential open text-to-speech models. The repository has about 7,900 stars on GitHub, is released under the MIT license, and became a standard baseline that countless later TTS projects build on. Its code was last updated in 2023.
Unlike classic two-stage pipelines that first predict a mel spectrogram and then hand it to a separate vocoder, VITS learns to generate raw waveforms end to end. It combines a conditional variational autoencoder with adversarial training and a stochastic duration predictor, which lets it produce natural, expressive speech with realistic variation in rhythm and intonation. The repository includes training and inference code for single-speaker synthesis on LJSpeech and multi-speaker synthesis on VCTK, together with pretrained checkpoints and an interactive demo. Because the training recipe is fully reproducible, the repo has served as the starting point for many community fine-tunes and language adaptations.
Free. The code is open source under the MIT license and the pretrained models can be downloaded and used at no cost.
Speech researchers and developers who want a proven, hackable end-to-end TTS model โ whether to study the architecture, fine-tune it on a new voice, or embed quality speech synthesis in a project.