Author:
Xie Xiaoping,Chen Yongzhen,Shen Rufeng,Tian Dan
Abstract
Abstract
Speech feature model is the basis of speech and noise separation, speech expression, and different styles of speech conversion. With the development of signal processing methods, the feature types and dimensions increase. Therefore, it is difficult to select appropriate features. If a single feature is used, the representation of the speech signal will be incomplete. If multiple features are used, there will be redundancy between features, which will affect the performance of speech separation. The feature described above is a combination of parameters to characterize speech. A single feature means that the combination has only one parameter. In this paper, the feature selection method is used to select and combine eight widely used speech features and parameters. The Deep Neural Network (DNN) is used to evaluate and analyze the speech separation effect of different feature groups. The comparison results show that the speech segregation effect of the complementary feature group is better. The effectiveness of the complementary feature group to improve the performance of DNN speech separation is verified.
Funder
Natural Science Foundation of Hunan Province
Publisher
Springer Science and Business Media LLC
Subject
Electrical and Electronic Engineering,Acoustics and Ultrasonics
Reference27 articles.
1. Z.X. Li, Research on Single-Channel Speech Separation Method Based on Autoregressive Deep Neural Network, University of Science and Technology of China (2019), pp.2–4
2. H. Li, Single-channel Speech Separation Based on Deep Learning, Inner Mongolia University (2017), pp.5–8
3. E. A. Hoer, Kennard, et al. Ridge Regression: Biased Estimation for Non-orthogonal Problems. Technometrics (1970), pp.55-59
4. R. Tibshirani, Regression shrinkage and selection via the LASSO. J. Roy. Stat. Soc. 73(1), 273–282 (1996)
5. B. Efron, T. Hastie, J.R. Tibshirani, Least angle regression. Ann. Stat. 32(2), 407–451 (2004)