Text Classification based on Associative Relational Networks for Multi-Domain Text-to-Speech Synthesis

Abstract

This work is a step further in our research towards developing a new strategy for high quality text-to-speech (TTS) synthesis among different domains. In this context, it is necessary to select the most appropriate domain for synthesizing the text input to the TTS system, task that can be solved including a text classifier (TC) in the classic TTS architecture. Since speech speaking style and prosody depend on the sequentiality and text structure of the message, the TC should consider not only thematic but also stylistic aspects of text. To this end, we introduce a new text modelling scheme based on an associative relational network, which represents texts as a weighted word-based graph. The conducted experiments validate the proposal in terms of both objective (text classification efficiency) and subjective (perceived synthetic speech quality) evaluation criteria.