Skip to content

StyleTTS 2

A TTS model that uses style diffusion and adversarial training

Visit official site

What StyleTTS 2 is

StyleTTS 2 is described as a text-to-speech model that uses large speech language models for style diffusion and adversarial training, with style modelled as a latent random variable to suit the text without reference speech.

What you can do with it

  • Generate a style that fits the text without a reference clip
  • Model style as a latent random variable with a diffusion model
  • Use large pretrained SLMs as discriminators
  • Train end to end with differentiable duration modelling

Model notes

  • The material states the model reaches human-level TTS synthesis
  • The source names WavLM among the pretrained SLMs used as discriminators
  • The source states it surpassed human recordings on single-speaker LJSpeech and matched them on multi-speaker VCTK
  • The source states LibriTTS training outperformed earlier public zero-shot extensions

Pros & cons

✓ What we like

  • No reference speech clip is needed to produce a fitting style
  • End-to-end training pairs a discriminator with duration modelling

! What to watch out for

  • Style output is generated by the diffusion model without a reference clip
  • Training relies on pretrained SLMs and differentiable duration modelling

FAQ

What is modelled as a latent random variable?

The material states style is modelled that way by a diffusion model to suit the text.

Which pretrained model is named as a discriminator?

The source names large pretrained speech language models such as WavLM.

Last reviewed: 2026-11-15

More AI audio tools tools

View all →

How we review