Hearer model based stress prediction for Chinese TTS system

Abstract

People often feel tired if they listen to synthesized speech for a long time. This is mainly because synthesized speech is too flat and never stresses the focus. Unlike traditional TTS research approaches of speaker simulation, the paper investigates stress prediction from the point of view of the hearer. An ideal hearer model is first proposed to predict the stress distribution based on the following hypothesis: people speak with limited stress effort and distribute the limited effort to ensure that the hearer can understand the speaker easily. Then, according to the limited research resource, we modify the ideal hearer model and present a practical model. Experiments show that the stress prediction achieves an acceptable rate of 87.36%.