关键词: 深度信念网络, 语音增强, 联合优化, 回归

Abstract: Concerning the problem that the poor generalization ability of Deep Believe Network（DBN） which leads to poor speech enhancement performance, a regression DBN speech enhancement algorithm based on features jointing optimization is proposed. It is not necessary to make any assumptions about speech and noise in advance. The Log-Mel frequency Power Spectrum（LMPS） of speech is extracted to be used directly for constructing the enhanced speech signals to ensure the quality of speech hearing, and the Mel-Frequency Cepstral Coefficients（MFCC） of speech is extracted as an auxiliary features, respectively. All the parameters of the original deep belief network architecture are optimized by integrating the combination feature into DBN system. This joint optimization estimation scheme imposes MFCC constraints not available in the direct prediction of LMPS, and improves the generalization ability of the model to estimate the LMPS, and reconstructs the enhanced speech more accurately. Simulation results in different SNR enviroment show that compared with single feature optimization such as Log Power Spectrum（LPS） and LMPS, LMPS and MFCC joint optimization can enable the enhanced speech obtain higher PESQ and SNR, and improve speech quality and intelligibility.

Key words: Deep Believe Network（DBN）, speech enhancement, joint optimization, regression