VITS: Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech

VITS is an end-to-end speech synthesis model that predicts a speech waveform conditional on an input text sequence. It is a conditional variational autoencoder (VAE) comprised of a posterior encoder, decoder, and conditional prior. This repository contains the weights for the official VITS checkpoint trained on the LJ Speech dataset.

Model Details

VITS (Variational Inference with adversarial learning for end-to-end Text-to-Speech) is an end-to-end speech synthesis model that predicts a speech waveform conditional on an input text sequence. It is a conditional variational autoencoder (VAE) comprised of a posterior encoder, decoder, and conditional prior.

A set of spectrogram-based acoustic features are predicted by the flow-based module, which is formed of a Transformer-based text encoder and multiple coupling layers. The spectrogram is decoded using a stack of transposed convolutional layers, much in the same style as the HiFi-GAN vocoder. Motivated by the one-to-many nature of the TTS problem, where the same text input can be spoken in multiple ways, the model also includes a stochastic duration predictor, which allows the model to synthesise speech with different rhythms from the same input text.

The model is trained end-to-end with a combination of losses derived from variational lower bound and adversarial training. To improve the expressiveness of the model, normalizing flows are applied to the conditional prior distribution. During inference, the text encodings are up-sampled based on the duration prediction module, and then mapped into the waveform using a cascade of the flow module and HiFi-GAN decoder. Due to the stochastic nature of the duration predictor, the model is non-deterministic, and thus requires a fixed seed to generate the same speech waveform.

There are two variants of the VITS model: one is trained on the LJ Speech dataset, and the other is trained on the VCTK dataset. LJ Speech dataset consists of 13,100 short audio clips of a single speaker with a total length of approximately 24 hours. The VCTK dataset consists of approximately 44,000 short audio clips uttered by 109 native English speakers with various accents. The total length of the audio clips is approximately 44 hours.

Usage

VITS is available in the 🤗 Transformers library from version 4.33 onwards. To use this checkpoint, first install the latest version of the library:

pip install --upgrade transformers accelerate

Then, run inference with the following code-snippet:

from transformers import VitsModel, AutoTokenizer
import torch

model = VitsModel.from_pretrained("kakao-enterprise/vits-ljs")
tokenizer = AutoTokenizer.from_pretrained("kakao-enterprise/vits-ljs")

text = "Hey, it's Hugging Face on the phone"
inputs = tokenizer(text, return_tensors="pt")

with torch.no_grad():
    output = model(**inputs).waveform

The resulting waveform can be saved as a .wav file:

import scipy

scipy.io.wavfile.write("techno.wav", rate=model.config.sampling_rate, data=output)

Or displayed in a Jupyter Notebook / Google Colab:

from IPython.display import Audio

Audio(output, rate=model.config.sampling_rate)

BibTex citation

This model was developed by Jaehyeon Kim et al. from Kakao Enterprise. If you use the model, consider citing the VITS paper:

@inproceedings{kim2021conditional,
  title={"Conditional Variational Autoencoder with Adversarial Learning for End-to-end Text-to-speech"},
  author={Kim, Jaehyeon and Kong, Jungil and Son, Juhee},
  booktitle={International Conference on Machine Learning},
  pages={5530--5540},
  year={2021},
  organization={PMLR}
}

License

The model is licensed as MIT.