Creating expressive synthetic voices by unsupervised clustering of audiobooks

Abstract

In this work we design an approach for automatic feature selection and voice creation for expressive synthesis. Our approach is guided by two main goals: (1) increasing the flexibility of expressive voice creation and (2) overcoming the limitations of speaking styles in expressive synthesis. We define a novel set of features, combining traditionally used prosodic features with spectral features and proposing the use of iVectors. With these features we perform unsupervised clustering of an audiobook excerpt and, from these clusters, we create synthetic voices using the SAT technique. To evaluate the clustering performance we propose an objective evaluation of the unsupervised clustering results technique based on perplexity reduction. This objective evaluation indicates that both prosodic and spectral features contribute to separate speaking styles and emotions, achieving the best results when including iVectors in the feature set, leading to a perplexity reduction of the expressions and audiobook characters by factors 14 and 2, respectively. We also designed a novel subjective evaluation method where the participants have to edit a small excerpt of an audiobook using synthetic voices created from clusters. The results suggest that our feature set is effective in the task of expressiveness and character detection.