StyleTTS 2

On the record

Code license

The repository states that its code is licensed under the MIT License.github.com · read 13 Oct 2026

Compute requirements

The repository notes that training or inference can run into out-of-memory issues and recommends reducing batch size or maximum audio length; it also says older GPUs may produce background noise.github.com · read 13 Oct 2026

Dependencies

The repository says its inference depends on a GPL-licensed package that is not included directly, and also links to a fully MIT-licensed package using gruut with lower output quality due to a phonemizer mismatch.github.com · read 13 Oct 2026

Evaluation

The makers report that StyleTTS 2 surpasses human recordings on single-speaker LJSpeech and matches them on multispeaker VCTK, as judged by native English speakers.styletts2.github.io · read 12 Oct 2026

Expressiveness

The sample page demonstrates expressive speech, style transfer, longform narration, and speech diversity.styletts2.github.io · read 13 Oct 2026

Inference

The repository provides inference notebooks and pretrained model downloads for LJSpeech and LibriTTS.github.com · read 13 Oct 2026

Installation

The repository lists Python 3.7 or newer as a prerequisite and gives additional installation instructions for Windows.github.com · read 13 Oct 2026

Known limitation

The repository says the distributed-data-parallel version of second-stage training does not work, so the current version uses data parallelism.github.com · read 13 Oct 2026

Language support

The repository says non-English training requires a pretrained PL-BERT model for the language and points to a multilingual PL-BERT supporting 14 languages.github.com · read 13 Oct 2026

Limitations

The README says the second-stage training code uses data parallelism because its distributed data parallel version does not work.github.com · read 12 Oct 2026

Linux dependency

The demo instructions say to install phonemizer and espeak-ng; the example espeak-ng command uses apt-get.github.com · read 12 Oct 2026

Model use

For pretrained models, the repository asks users to disclose synthesized speech unless they have permission to use the cloned voice, and says public voice use requires speaker permission or disclosure.github.com · read 13 Oct 2026

Online demo

The repository links to a Hugging Face online demo.github.com · read 13 Oct 2026

Open source license

The repository code is licensed under the MIT License.github.com · read 12 Oct 2026

Pretrained models

The repository links pretrained models for LJSpeech and LibriTTS.github.com · read 12 Oct 2026

Purpose

StyleTTS 2 is a text-to-speech model that uses style diffusion and adversarial training with large speech language models.styletts2.github.io · read 13 Oct 2026

Speaker adaptation

The project reports that its LibriTTS-trained model outperforms previous publicly available models for zero-shot speaker adaptation.styletts2.github.io · read 13 Oct 2026

Speech naturalness

The model uses pre-trained speech language models such as WavLM as discriminators and differentiable duration modeling to improve speech naturalness.styletts2.github.io · read 12 Oct 2026

Alternatives to StyleTTS 2

GC Google Cloud Text-to-SpeechAndroidiPhone & iPad+2 V VoiceboxAndroidiPhone & iPad+6 RA Resemble AIAndroidiPhone & iPad+7 T TypecastAndroidiPhone & iPad+2 S SpeechifyAndroidiPhone & iPad+5 N NaturalReaderAndroidiPhone & iPad+4 T TTSReaderAndroidiPhone & iPad+5 I iSpeechAndroidiPhone & iPad+6

Every alternative, ranked