StyleTTS2

StyleTTS2 is the research TTS model that claimed the first human-level synthesis on LJSpeech and VCTK: it models speaking style as a latent random variable via diffusion (no reference speech needed at inference) and trains adversarially against large pretrained speech language models like WavLM with differentiable duration modeling. Pretrained single-speaker (LJSpeech) and multi-speaker (LibriTTS) checkpoints are on Hugging Face with Colab demos, plus a finetuning script for new speakers and strong zero-shot speaker adaptation. An independent 2025 human-fooling-rate study ranked it the best open-source TTS system, ahead of XTTS, F5-TTS, and GPT-SoVITS — though below actual human recordings. MIT licensed, but note the GPL dependency caveat and the voice-consent rule on pretrained models.

Research-repo path (verified from README):

```bash

git clone https://github.com/yl4579/StyleTTS2.git && cd StyleTTS2

pip install -r requirements.txt

pip install phonemizer && sudo apt-get install espeak-ng

```

Then run the demo notebooks: Demo/Inference_LJSpeech.ipynb (single-speaker) or Demo/Inference_LibriTTS.ipynb (multi-speaker; needs reference_audio.zip from Hugging Face yl4579/StyleTTS2-LibriTTS). Checkpoints: https://huggingface.co/yl4579/StyleTTS2-LJSpeech and https://huggingface.co/yl4579/StyleTTS2-LibriTTS.

Packaged path: pip install styletts2 (fully MIT-licensed, uses gruut — README notes lower quality than the phonemizer path); GPL-licensed importable fork with streaming API at github.com/NeuralVox/StyleTTS2.

Licensing note: the README requires you to disclose that outputs are synthesized (or have the speaker's permission) when using pretrained models on non-training-set voices.