VITS TTS

The end-to-end TTS architecture that combines a variational autoencoder, normalizing flows, and adversarial training.

VITS — Variational Inference with adversarial learning for end-to-end Text-to-Speech — was introduced by Jaehyeon Kim and collaborators in 2021 and became a foundational architecture for modern neural speech. Rather than the older two-stage pipeline, it synthesizes audio in a single parallel end-to-end pass, pairing a variational autoencoder with normalizing flows and a GAN-style adversarial training process to lift naturalness. At about 25M parameters and trained on ~585 hours, it produces natural prosody at fast inference speeds and supports multiple speakers. It serves as a solid general-purpose, free baseline and underpins many later models such as Piper and MeloTTS.