Audio Samples from Unofficial Implementation of UnivNet vocoder

Paper: Jang et al., UnivNet: A Neural Vocoder with Multi-Resolution Spectrogram Discriminators for High-Fidelity Waveform Generation

Implementation authors: Kang-wook Kim, Wonbin Jung @MINDsLab Inc.

GitHub repository: mindslab-ai/univnet

Please check the Github repository for the implementation details and the pre-trained model checkpoint.

The ground truth audio samples from the official demo page have a different volume from the original dataset. We believe the official ground truth audio samples have been post-processed. Therefore, the audio volume of our samples might have a different volume from the official version.

Notes: Both UnivNet-c16 and c32 results and the pre-trained weights have been uploaded. Our implementation matches the objective scores (PESQ and RMSE) of the original paper.


Synthesis for seen speakers

  • LibriTTS/train-clean-360 dataset was used to both train and evaluate the model.
  • The first five texts are taken directly from the official demo page for comparison.

  • Text

    Nobody could touch the body until the coroner came.

    Then they came to another grove of trees, where all the leaves were of gold; and afterwards to a third, where the leaves were all glittering diamonds.

    The shaven face of the priest is a further item to the same effect.

    The lady who had spoken English approached the table as if looking for something, and when Beaton looked again, the portrait had been turned on its face.

    He shouted, 'Where art thou, ring?' And the ring said, 'I am here,' though it was on the bed of the ocean.

    The monarch, after making some inquiry into the rank and character of his rival, despatched the informer with a present of a pair of purple slippers, to complete the magnificence of his Imperial habit.

    GT
    Official UnivNet-c16
    -
    Official UnivNet-c32
    -
    Our UnivNet-c16
    Our UnivNet-c32

    Synthesis for unseen speakers

  • LibriTTS/test-clean dataset was used to evaluate the model.
  • The first five texts are taken directly from the official demo page for comparison.

  • Text

    The stars began to crumble and a cloud of fine stardust fell through space.

    On August 27, 1837, she writes:--

    But when his big brother heard that he had refused to give his cap for a King's golden crown, he said that Anders was a stupid.

    The only thing is, I don't know anything about technique and stagecraft and the three unities and that sort of rot.

    The marquis of Worcester, a man past eighty-four, was the last in England that submitted to the authority of the parliament.

    True history being a mixture of all things, the true historian mingles in everything.

    GT
    Official UnivNet-c16
    -
    Official UnivNet-c32
    -
    Our UnivNet-c16
    Our UnivNet-c32