Effective Data Augmentation Methods for Neural Text-to-Speech Systems

This paper proposes an effective self-augmentation method for improving the quality of neural text-to-speech (TTS) systems. As synthetic speech quality has been greatly improved, creating a neural TTS system using synthetic corpora is now possible. However, whether increasing the amount of synthetic data is always beneficial for improving training efficiency has not been verified. Our aim in this study is to selectively choose synthetic data whose characteristics are close to those of natural speech. Specifically, we adopt a ranking support vector machine (RankSVM) that is well known for effectively ranking relative attributes among binary classes. By setting the synthetic and recorded corpora as two opposite classes, RankSVM is used to determine how the synthesized speech is acoustically similar with the recorded data. As training data can be selectively chosen from large-scale synthetic corpora, the performance of the TTS model re-trained by those data is significantly improved. Subjective evaluation results verify that the proposed TTS model performs much better than the original model trained with recorded data alone and the similarly configured system re-trained with all the synthetic data without any selection method.