Convolutional Pitch Target Approximation Model for Speech Synthesis

In this paper, we investigate pitch contour modelling in speech synthesis based on segmental units. A convolutional pitch target approximation model is proposed. This model allows jointly stochastic modelling of framewise pitch and pitch contour of longer units, of which the intuitive relations are revealed by a convolutional target approximation filter. The pitch contour is stylized by a linear representation called pitch target. In synthesis stage, the likelihood of the framewise model and the pitch target model are jointly maximized using a Toeplitz matrix representing the discrete convolutional filter.

Related material