Evaluation of naturalnEss of synthEsizEd spEEch with diffErEnt prosodic modEls

Obtaining natural synthesized speech is the main goal of modern research in the field of speech synthesis. It strongly depends on the prosody model used in the text-to-speech (TTS) system. This paper deals with speech synthesis evaluation with respect to the prosodic model used. Our Russian VitalVoice TTS is a unit selection concatenative system. We describe two approaches to prosody prediction used in VitalVoice Russian TTS. These are a rule-based approach and a hidden Markov model (HMM) based hybrid approach. We conduct an experiment for evaluating the naturalness of synthesized speech. Four variants of synthesized speech depending on the applied approach and the speech corpus size were tested. We also included natural speech samples into the test. Subjects had to rate the samples from 0 to 5 depending on their naturalness. The experiment shows that speech synthesized using the hybrid HMM-based approach sounds more natural than other synthetic variants. We discuss the results and the ways for further investigation and improvements in the last section.