关键词: 自编码特征, 深度神经网络, 特征提取, 信噪比

Abstract: In speech enhancement with supervised learning, feature extraction is a key step. Auditory features such as existing combined features and MRCG (multi-resolution cochleagram) are commonly used. Although the intelligibility of the enhanced speech based on these features is greatly improved, there is still a lot of noise in the enhanced speech and the quality (expressed as SNR) is low. In order to improve the quality after speech enhancement, without affecting the intelligibility, the IF (integrated feature) based on AEF (auto-encoder feature) is proposed. Firstly, the AEF is extracted by auto-encoder. Then, Group Lasso algorithm is used to verify the complementarity and redundancy between the AEF and the auditory features. These features are recombined to form IF. Finally, IF is used as input feature of speech enhancement system. Experiments are carried out on TIMIT corpus and Noisex-92 noise libraries. Compared with traditional speech enhancement methods, as well as deep learning methods using the existing combined features and MRCG as input features of speech enhancement system, the experimental results show that the speech effect of the proposed algorithm is greatly improved.

Key words: auto-encoder feature, deep neural network, feature extraction, signal noise ratio (SNR)