Unsupervised learning for expressive speech synthesis

Resumen

Nowadays, especially with the upswing of neural networks, speech synthesis is almost totally data driven. The goal of this thesis is to provide methods for automatic and unsupervised learning from data for expressive speech synthesis. In comparison to ¿ordinary¿ synthesis systems, it is more difficult to find reliable expressive training data, despite huge availability on sources like Internet. The main difficulty consists in the highly speaker- and situation-dependent nature of expressiveness, causing many and acoustically substantial variations. The consequences are, first, it is very difficult to define labels which reliably identify expressive speech with all nuances. The typical definition of 6 basic emotions, or alike, is a simplification which will have inexcusable consequences dealing with data outside the lab. Second, even if a label set is defined, apart of the enormous manual effort, it is difficult to gain sufficient training data for the models respecting all the nuances and variations. The goal of this thesis is to study automatic training methods for expressive speech synthesis avoiding labeling and to develop applications from these proposals. The focus lies on the acoustic and the semantic domains. For the part of the acoustic domain, the goal is to find suitable acoustic features to represent expressive speech, especially for the multi-speaker domain, as getting closer to real-life uncontrolled data. For this, the perspective will slide away from traditional, mainly prosody-based, features towards features gained with factor analysis, trying to identify the principal components of the expressiveness, namely using i-vectors. Results show that a combination of traditional and i-vector based features performs better in unsupervised clustering of expressive speech than traditional features and even better than large state-of-the-art sets in the multi-speaker domain. Once the feature set is defined, it is used for unsupervised clustering of an audiobook, where from each cluster a voice is trained. Then, the method is evaluated in an audiobook-editing application, where users can use the synthetic voices to create their own dialogues. The obtained results validate the proposal. In this editing application users choose synthetic voices and assign them to the sentences considering the speaking characters and the expressiveness. Involving the semantic domain, this assignment can be achieved automatically, at least partly. Words and sentences are represented numerically in trainable semantic vector spaces, called embeddings, and these can be used to predict the expressiveness to some extent. This method not only permits fully automatic reading of larger text passages, considering the local context, but can also be used as a semantic search engine for training data. Both applications are evaluated in a perceptual test showing the potential of the proposed method. Finally, accounting for the new tendencies in the speech synthesis world, deep neural network based expressive speech synthesis is designed and tested. Emotionally motivated semantic representations of text, sentiment embeddings, trained on the positiveness and the negativeness of movie reviews, are used as an additional input to the system. The neural network now learns not only from segmental and contextual information, but also from the sentiment embeddings, affecting especially prosody. The system is evaluated in two perceptual experiments which show preferences for the inclusion of sentiment embeddings as an additional input.