Reproduce “Statistical Parametric Speech Synthesis Incorporating Generative Adversarial Networks” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yuki Saito, Shinnosuke Takamichi, Hiroshi Saruwatari
Researchers at The University of Tokyo integrated Generative Adversarial Networks (GANs) into statistical parametric speech synthesis to reduce the "over-smoothing effect." This approach produced synthetic speech with distributions closer to natural speech and achieved significant perceptual quality improvements in both text-to-speech and voice conversion tasks.
Key insights
- GAN training effectively mitigates the over-smoothing effect in statistical parametric speech synthesis by restoring natural variance and correlation patterns in generated acoustic parameters.
- Among various GAN formulations, Wasserstein GAN (W-GAN) and Least Squares GAN (LS-GAN) proved most effective for improving speech quality in this context.
- While beneficial for spectral and F0 parameters, adversarial training showed limited subjective improvement for duration modeling, suggesting task-specific discriminator designs or objectives may be required for temporal features.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.