Estimation of Pitch Targets from Speech Signals by Joint Regularized Optimization

Abstract

This paper presents a novel method to estimate the pitch target parameters of the target approximation model (TAM). The TAM allows the compact representation of natural pitch contours on a solid theoretical basis and can be used as an intonation model for text-to-speech synthesis. In contrast to previous approaches, the method proposed here estimates the parameters of all targets jointly, uses 5th-order (instead of 3rd-order) linear systems to model the target approximation process, and uses regularization to avoid unnatural pitch targets. The effect of these features on the modeling error and the target parameter distributions are shown. The proposed method has been made available as the open-source software tool TargetOptimizer.