中国科学院自动化研究所机构知识库

The fundamental frequency and the spectrum of the speech are related thus one of their learned mapping from the linguistic features can be leveraged for another. The conventional methods treat all the acoustic features as one stream for acoustic modeling. And the multi-task learning methods are applied to train the model simultaneously with serval targets and a combined cost. To improve the accuracy of the acoustic model, the progressive deep neural networks (PDNN) is applied for acoustic modeling in statistical parametric speech synthesis (SPSS) in our method. Each type of the acoustic features is modeled in different networks with its own cost function and the knowledge transfers through lateral connections. Each networks in the progressive networks can be trained to reach its own optimal step by step. Experiments are conducted to compare the PDNN based SPSS and the DNN based SPSS. The DNN and PDNN with the multi-task learning (MTL) method for acoustic modeling are also the tested. The computational complexity, prediction sequences and quantity of hierarchies of the PDNN are investigated. Both objective and subjective experimental results demonstrate the effectiveness of the proposed technique.