Learning Prosodic Stress from Data in Neural Network based Text-to-Speech Synthesis

Abstract

Naturalness is one of the most important aspects of synthesized speech, and state-of-the-art parametric speech synthesizers require training on large quantities of annotated speech data to be able to convey prosodic elements such as pitch accent and phrase boundary tone. The most frequently used framework for prosodic annotation of speech in American English is Tones and Break Indices – ToBI, which has also been adapted for use in a number of other languages. This paper presents certain deficiencies of ToBI when applied in synthesis of speech in American English, which are related to the absence of tags specifically intended to mark differences in the level of prosodic stress (emphasis) related to a particular sentence constituent. The research presented in the paper proposes the introduction of a set of tags intended for explicit modeling of the degree of prosodic stress. Namely, a certain sentence constituent can be particularly emphasized, when it is the intended focus of the utterance, or it can be de-emphasized, as is commonly the case with phrases reporting direct speech or with comment clauses. Through several listening tests it has been shown that learning such prosodic events from data has distinct advantages over approaches attempting to exploit the existing ToBI tags to convey the degree of emphasis in synthesized speech. Namely, speech synthesized by a neural network trained on data tagged for the level of prosodic stress appears more natural, and the listeners are more successful in locating the sentence constituent carrying prosodic stress.