VITS

VITS (Variational Inference with Adversarial Learning for End-to-End Text-to-Speech) was proposed by Jaehyeon Kim (jaywalnut310) and colleagues at Seoul National University, and is a landmark work in the field of speech synthesis models. VITS was the first to introduce an end-to-end architecture combining variational autoencoders (VAE) with adversarial training, achieving single-stage synthesis directly from text to waveform — bypassing the cumbersome intermediate representations of traditional two-stage approaches (acoustic model + vocoder). This design enabled VITS to achieve state-of-the-art performance in both synthesis speed and quality at the time of its publication at ICML 2021.

Architecturally, VITS’s core is a conditional VAE comprising a posterior encoder, prior encoder, flow-based decoder, and an adversarial discriminator combined with HiFi-GAN. During training, the model is jointly optimized using a variational lower bound and an adversarial loss; during inference, only a single forward pass is needed to generate high-fidelity audio. VITS natively supports multi-speaker training and multilingual datasets, achieving synthesis quality significantly superior to traditional combinations such as Tacotron 2 + WaveGlow under equivalent training conditions. The project is licensed under MIT, with clean and concise code that has been widely regarded as a best-practice reference for understanding end-to-end open-source TTS architecture since its release.

VITS’s influence extends far beyond the project itself — many subsequent popular speech synthesis model projects (including VITS2, Bert-VITS2, and components of GPT-SoVITS) have built upon and extended its architecture. The project’s PyTorch implementation, detailed training instructions, and preprocessing scripts enable researchers and developers to quickly reproduce results and adapt the model to new languages and scenarios. Although VITS does not include explicit voice cloning or emotion control capabilities, the end-to-end paradigm it established has profoundly shaped the direction of TTS research and remains one of the most frequently used foundational TTS architectures on GitHub.