StyleTTS 2 is an open-source text-to-speech model that reaches human-level quality through style diffusion and adversarial training with large speech language models. The repository has about 6,400 stars on GitHub, with the latest code updates in 2024.
StyleTTS 2 models speaking style with a diffusion process and refines the result through adversarial training against a large speech language model built on WavLM. In the project’s evaluations, it surpassed human recordings on the LJSpeech dataset and matched human quality on the multi-speaker VCTK corpus — results that made it one of the reference points for open TTS research. The model supports zero-shot speaker adaptation, so it can adopt a new voice from a short reference clip without retraining. Pretrained models are published on Hugging Face, and the codebase is the official implementation of the StyleTTS 2 paper, which keeps it popular with researchers who need a strong, reproducible baseline. The repository includes inference code plus links to demos and pretrained checkpoints, so the published results can be reproduced without training from scratch. While newer TTS systems have appeared since, StyleTTS 2 remains a widely cited milestone: its combination of style diffusion and adversarial training showed how close open models could come to studio recordings.
The code is free and open source under the MIT license. The pretrained models come with the expectation that listeners are informed the speech is AI-generated.
Speech researchers and developers who need a high-quality open TTS baseline for narration, accessibility, prototyping, or as a foundation for fine-tuning.