vits

Why we included this project

VITS strings a whole text-to-speech pipeline into a single model, so you skip the usual chain of separate acoustic models and vocoders. It's built around a conditional variational autoencoder with normalizing flows and adversarial training, and the stochastic duration predictor is what gives the voice variety: the same sentence can come out with different pitch and rhythm instead of one fixed cadence. The authors backed the code with a published paper, pretrained models for the LJ Speech single-speaker dataset, and an interactive Colab demo, which means you can judge the audio before touching any config. Training and inference scripts cover both single-speaker and multi-speaker (VCTK) setups, so it works as a reproducible baseline for research or a base to adapt to your own data.