SC-GlowTTS

Abstract:

In this paper, we propose SC-GlowTTS: an efficient zero-shot multi-speaker text-to-speech model that improves similarity for speakers unseen in training. We propose a speaker conditional architecture that explores a flow-based decoder which is able to work in a zero-shot scenario. As text encoders, we explored a dilated residual convolutional-based encoder, gated convolutional-based encoder and transformer-based encoder. Additionally, we have shown that adjusting a GAN-based vocoder for the spectrograms predicted by the TTS model on the training dataset can significantly improve the similarity and speech quality for new speakers. We showed that our model can converge in training, using only 11 speakers, reaching state-of-the-art results for similarity with new speakers and speech quality.

System architecture:

Audio Samples for VCTK test sentences

Trained using 98 speakers:

Trained using 11 speakers:

Audio Samples MOS sentences (LibriTTS test clean)

Zero-Shot Voice Conversion

This does not use HiFi-GAN-FT so the quality depends a lot on the reference audio, feel free to adjust the model

The rows present the ground truth audios converted into the voice of the speakers of each column.

Citation

        @inproceedings{casanova21b_interspeech,
	  author={Edresson Casanova and Christopher Shulby and Eren Gölge and Nicolas Michael Müller and Frederico Santos de Oliveira and Arnaldo {Candido Jr.} and Anderson da Silva Soares and Sandra Maria Aluisio and Moacir Antonelli Ponti},
	  title={{SC-GlowTTS: An Efficient Zero-Shot Multi-Speaker Text-To-Speech Model}},
	  year=2021,
	  booktitle={Proc. Interspeech 2021},
	  pages={3645--3649},
	  doi={10.21437/Interspeech.2021-1774}
	}