Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Exploring synthetic data for cross-speaker style transfer in style representation based TTS

Lucas H. Ueda, Leonardo B. de M. M. Marques, Flávio O. Simões, Mário U. Neto, Fernando Runstein, Bianca Dal Bó, Paula D. P. Costa

TL;DR

The results show that using VC synthetic data can improve the naturalness and speaker similarity of TTS in cross-speaker scenarios and extend this approach to a cross-language scenario, enhancing accent transfer.

Abstract

Incorporating cross-speaker style transfer in text-to-speech (TTS) models is challenging due to the need to disentangle speaker and style information in audio. In low-resource expressive data scenarios, voice conversion (VC) can generate expressive speech for target speakers, which can then be used to train the TTS model. However, the quality and style transfer ability of the VC model are crucial for the overall TTS model quality. In this work, we explore the use of synthetic data generated by a VC model to assist the TTS model in cross-speaker style transfer tasks. Additionally, we employ pre-training of the style encoder using timbre perturbation and prototypical angular loss to mitigate speaker leakage. Our results show that using VC synthetic data can improve the naturalness and speaker similarity of TTS in cross-speaker scenarios. Furthermore, we extend this approach to a cross-language scenario, enhancing accent transfer.

Exploring synthetic data for cross-speaker style transfer in style representation based TTS

TL;DR

The results show that using VC synthetic data can improve the naturalness and speaker similarity of TTS in cross-speaker scenarios and extend this approach to a cross-language scenario, enhancing accent transfer.

Abstract

Incorporating cross-speaker style transfer in text-to-speech (TTS) models is challenging due to the need to disentangle speaker and style information in audio. In low-resource expressive data scenarios, voice conversion (VC) can generate expressive speech for target speakers, which can then be used to train the TTS model. However, the quality and style transfer ability of the VC model are crucial for the overall TTS model quality. In this work, we explore the use of synthetic data generated by a VC model to assist the TTS model in cross-speaker style transfer tasks. Additionally, we employ pre-training of the style encoder using timbre perturbation and prototypical angular loss to mitigate speaker leakage. Our results show that using VC synthetic data can improve the naturalness and speaker similarity of TTS in cross-speaker scenarios. Furthermore, we extend this approach to a cross-language scenario, enhancing accent transfer.
Paper Structure (11 sections, 2 figures, 5 tables)

This paper contains 11 sections, 2 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Synthetic data
  4. Stage 1: Style Encoder
  5. Stage 2: TTS
  6. Experiments
  7. Training setup
  8. Evaluation
  9. Results
  10. Conclusion and future work
  11. Acknowledgements

Figures (2)

  • Figure 1: Two stages TTS training pipeline. Style encoder is pre-trained in stage 1 and remains frozen during TTS training at stage 2. The speaker look-up table encodes speaker information in FastPitch for multi-speaker extension.
  • Figure 2: Style representations projected using UMAP McInnes2018 when training with and without synthetic data in Stage 1 ("x" markers are synthetic expressive data).