Comparative Evaluation of Neural Vocoders for Speech Synthesis of Operatic Singing

Sota Shimizu, Keisuke Matsubara, Yuji Adachi, Kiyoto Tai, Ryoichi Takashima, Tetsuya Takiguchi

Abstract: The voice of someone who is singing opera has different characteristics compared with someone singing nursery rhymes or pop songs, which are treated by conventional singing voice synthesis. In this study, we compared and evaluated the speech synthesis performance of recently proposed neural vocoders for an operatic singing voice. From the results of our objective evaluation experiments, it was found that the melcepstrum estimation error for Parallel WaveGAN and PeriodNet is small. Furthermore, PeriodNet has a smaller estimation error of fundamental frequency.