Control of fundamental frequency contours using the generation process model in HMM-based speech synthesis

Abstract

A method was proposed to increase the naturalness of prosody generated with speech synthesis based on hidden Markov models (HMMs). This method adds a constraint to the fundamental frequency contours (F <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">0</sub> contours) during the HMM-based speech synthesis. The constraint adopted is the generation process model of F <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">0</sub> contours (F <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">0</sub> model). The method first extracts the F <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">0</sub> model parameters from the original Fo contour (generated by the HMM-based speech synthesis) and then optimizes them successively by referring to the pre-trained HMMs. The experimental results show that the proposed method can improve naturalness when the F <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">0</sub> control by the original method is inadequate.