Statistical Parametric Speech Synthesis Based on Gaussian Process Regression

This paper proposes a statistical parametric speech synthesis technique based on Gaussian process regression (GPR). The GPR model is designed for directly predicting frame-level acoustic features from corresponding information on frame context that is obtained from linguistic information. The frame context includes the relative position of the current frame within the phone and articulatory information and is used as the explanatory variable in GPR. Here, we introduce cluster-based sparse Gaussian processes (GPs), i.e., local GPs and partially independent conditional (PIC) approximation, to reduce the computational cost. The experimental results for both isolated phone synthesis and full-sentence continuous speech synthesis revealed that the proposed GPR-based technique without dynamic features slightly outperformed the conventional hidden Markov model (HMM)-based speech synthesis using minimum generation error training with dynamic features.