Abstract:

Recently, text-to-speech (TTS) synthesis has gained immense success in the human-computer interaction domain. Current TTS systems are monotonous due to the absence of expressivity. Expressivity in speech generally refers to suprasegmental speech characteristics represented by emotions, speaking styles, and the relationship between speech and gestures, facial expressions, etc. It seems likely that expressive speech synthesis provides the ability to improve the user experience with machines greatly. The development of an expressive TTS system heavily relies on the speech data used in training the system. The thesis aims at developing an expressive TTS system in a speaker’s voice for which only neutral speech data is available. The main focus of the thesis is to investigate deep learning approaches for exploring the disentanglement of speaker information and expressivity in a multispeaker TTS setting. The scope of the work incorporates expressivity as an emotion attribute with well-defined emotion classes.

Jury :

Rapporteures :

Yannis Stylianou Professeur – University of Crete

Damien Lolive Maitre de Conférences HDR – Université de Rennes 1

Examinateurs :

Marie Tahon Maitre de Conférences – Université du Mans

Axel Roebel Directeur de Recherches IRCAM

Directrices de thèse :

Denis Jouvet Directeur de Recherches Inria

Vincent Colotte Maitre de Conf ́erences – Universit ́e de Lorraine