TY - GEN
T1 - Statistical parametric speech synthesis using generative adversarial networks under a multi-task learning framework
AU - Yang, Shan
AU - Xie, Lei
AU - Chen, Xiao
AU - Lou, Xiaoyan
AU - Zhu, Xuan
AU - Huang, Dongyan
AU - Li, Haizhou
N1 - Publisher Copyright:
© 2017 IEEE.
PY - 2017/7/2
Y1 - 2017/7/2
N2 - In this paper, we aim at improving the performance of synthesized speech in statistical parametric speech synthesis (SPSS) based on a generative adversarial network (GAN). In particular, we propose a novel architecture combining the traditional acoustic loss function and the GAN's discriminative loss under a multi-task learning (MTL) framework. The mean squared error (MSE) is usually used to estimate the parameters of deep neural networks, which only considers the numerical difference between the raw audio and the synthesized one. To mitigate this problem, we introduce the GAN as a second task to determine if the input is a natural speech with specific conditions. In this MTL framework, the MSE optimization improves the stability of GAN, and at the same time GAN produces samples with a distribution closer to natural speech. Listening tests show that the multi-task architecture can generate more natural speech that satisfies human perception than the conventional methods.
AB - In this paper, we aim at improving the performance of synthesized speech in statistical parametric speech synthesis (SPSS) based on a generative adversarial network (GAN). In particular, we propose a novel architecture combining the traditional acoustic loss function and the GAN's discriminative loss under a multi-task learning (MTL) framework. The mean squared error (MSE) is usually used to estimate the parameters of deep neural networks, which only considers the numerical difference between the raw audio and the synthesized one. To mitigate this problem, we introduce the GAN as a second task to determine if the input is a natural speech with specific conditions. In this MTL framework, the MSE optimization improves the stability of GAN, and at the same time GAN produces samples with a distribution closer to natural speech. Listening tests show that the multi-task architecture can generate more natural speech that satisfies human perception than the conventional methods.
KW - conditional generative adversarial network
KW - deep neural network
KW - generative adversarial network
KW - multi-task learning
KW - Statistical parametric speech synthesis
UR - http://www.scopus.com/inward/record.url?scp=85047518812&partnerID=8YFLogxK
U2 - 10.1109/ASRU.2017.8269003
DO - 10.1109/ASRU.2017.8269003
M3 - 会议稿件
AN - SCOPUS:85047518812
T3 - 2017 IEEE Automatic Speech Recognition and Understanding Workshop, ASRU 2017 - Proceedings
SP - 685
EP - 691
BT - 2017 IEEE Automatic Speech Recognition and Understanding Workshop, ASRU 2017 - Proceedings
PB - Institute of Electrical and Electronics Engineers Inc.
T2 - 2017 IEEE Automatic Speech Recognition and Understanding Workshop, ASRU 2017
Y2 - 16 December 2017 through 20 December 2017
ER -