Sequential Generation of Singing F0 Contours from Musical Note Sequences Based on WaveNet

This paper describes a method that can generate a continuous F0 contour of a singing voice from a monophonic sequence of musical notes (musical score) by using a deep neural autoregressive model called WaveNet. Real F0 contours include complicated temporal and frequency fluctuations caused by singing expressions such as vibrato and portamento. Although explicit models such as hidden Markov models (HMMs) have often used for representing the F0 dynamics, it is difficult to generate realistic F0 contours due to the poor representation capability of such models. To overcome this limitation, WaveNet, which was invented for modeling raw waveforms in an unsupervised manner, was recently used for generating singing F0 contours from a musical score with lyrics in a supervised manner. Inspired by this attempt, we investigate the capability of WaveNet for generating singing F0 contours without using lyric information. Our method conditions WaveNet on pitch and contextual features of a musical score. As a loss function that is more suitable for generating F0 contours, we adopted the modified cross-entropy loss weighted with the square error between target and output F0s on the log-frequency axis. The experimental results show that these techniques improve the quality of generated F0 contours.