Automatic Exploration of Corpus-Specific Properties for Expressive Text-to-Speech: A Case Study in Emphasis

Abstract

In this paper we explore an approach to expressive text-to-speech synthesis in which pre-existing expression-specific corpora are complemented with automatically generated labels to augment the search space of units the engine can exploit to increase its expressiveness. We motivate this data-discovery approach as an alternative to an approach guided by data collection, in order to harness the full usefulness of the expressiveness already contained in a synthesis corpus. We illustrate the approach with a case study that uses emphasis as its intended expression, describe algorithms for the automatic discovery of such instances in the database and how to make use of them during synthesis, and, finally, evaluate the benefits of the proposal to demonstrate the feasibility of the approach.