Using Audio Books for Training a Text-to-Speech System

Creating new voices for a TTS system often requires a costly procedure of designing and recording an audio corpus, a time consuming and effort intensive task. Using publicly available audiobooks as the raw material of a spoken corpus for such systems creates new perspectives regarding the possibility of creating new synthetic voices quickly and with limited effort. This paper addresses the issue of creating new synthetic voices based on audiobook data in an automated method. As an audiobook includes several types of speech, such as narration, character playing etc., special care is given in identifying the data subset that leads to a more neutral and general purpose synthetic voice. Part of the work described in this paper was carried out for the participation of our TTS system in the Blizzard Challenge 2013 where the developed TTS system was ranked top in overall impression in one of the two experimental hubs (Chalamandaris et al 2013). The main goal is to identify and address the effect the audiobook speech diversity on the resulting TTS system. Along with the methodology for coping with this diversity in the speech data, we also describe a set of experiments performed in order to investigate the efficiency of different approaches for automatic data pruning. Further plans for exploiting the diversity of the speech incorporated in an audiobook are also described in the final section and conclusions are drawn.