Abstract
Recent researches on both utterance-level and phone-level prosody modelling successfully improve the voice quality and naturalness in text-to-speech synthesis. However, most of them model the prosody with a unimodal distribution such like a single Gaussian, which is not reasonable enough. In this work, we focus on phone-level prosody modelling where we introduce a Gaussian mixture model(GMM) based mixture density network. Our experiments on the LJSpeech dataset demonstrate that GMM can better model the phone-level prosody than a single Gaussian. The subjective evaluations suggest that our method not only significantly improves the prosody diversity in synthetic speech without the need of manual control, but also achieves a better naturalness. We also find that using the additional mixture density network has only very limited influence on inference speed.
Abstract (translated)
URL
https://arxiv.org/abs/2102.00851
