Progressive Neural Networks based Features Prediction for the Target Cost in Unit-Selection Speech Synthesizer

This paper describes a direct acoustic features prediction for calculation of the target cost by progressive neural networks. Compared with conventional methods involving many hand-tuning steps, our method directly predicts the features for calculation of the target cost. By applying the progressive deep neural network (PDNN) to predict these acoustic features, the correlation of these features can be modeled. Each type of the acoustic features and each part of a unit are modeled in different sub-networks with its own cost function and the knowledge transfers through lateral connections. Each sub-network in the PDNN can be trained to reach its own optimum step by step. Extensive comparative evaluations demonstrate the effectiveness of the PDNN in improving the accuracy of predicted acoustic features. The subjective evaluation results demonstrate that the naturalness of synthetic speech has been improved by adopting the proposed method to calculate the target cost.