HMM-based Speech Synthesis from Audio Book Data

Abstract

In contrast to hand-crafted speech databases, which contain short out-of-context sentences in fairly unemphatic speech style, audio books contain rich prosody including intonation contours, pitch accents and phrasing patterns, which is a good pre-requisite for building a natural sounding synthetic voice. The following paper will give an overview of the steps that are involved in building a synthetic voice from audio book data. After an introduction to the theory of HMM-based speech synthesis, the properties of the speech database will be described in detail. It will be argued that it is necessary to model specific properties of the database, such as higher pitched speech or questions, to achieve a better quality synthetic voice. Furthermore, the acoustic modelling of these properties will be explained in detail. Finally, the synthetic voice is evaluated on the basis of an online listening test.