A High Intelligibility Signal Subspace Speech-enhancement Algorithm

Nowadays speech communication has become an essential part of our modern life. However, with noise pollution increasing, the SNR (Signal-to-Noise Ratio) of the noisy speech that people receive is getting lower and lower, which seriously interferes with the normal understanding of the speech. Study of the existing speech enhancement algorithms mainly focuses on how to improve the quality of the speech, while the speech intelligibility especially how to improve the intelligibility of the speech in low SNR environments is relatively less studied. To compensate for this deficiency, this paper makes a summing up of signal subspace speech enhancement algorithm. Then it makes an analysis of the reasons why the existing speech enhancement algorithms cannot improve speech intelligibility. What’s more, from the perspective that different types of speech distortions have different effects on the intelligibility of the enhanced speech, a high intelligibility subspace algorithm is proposed in this paper which is based on the classification of speech distortions in order to improve effectively the intelligibility of noisy speech in low SNR environments.The main research work is embodied in the following three aspects:First, we summarized the basic knowledge of speech enhancement and classified and analyzed the common algorithms of speech enhancement. Speech enhancement has been a research focus and an important branch of speech processing technology. Therefore, we first made an analysis of speech characteristics, perceptual characteristics of the human ear and the characteristics of the noise. Then we established a speech-enhanced signal model and summarized the short-time preprocessing common methods in the speech processing. The existing speech enhancement algorithms are complex. Therefore we classified and sorted these algorithms, and made a comparative analysis of the advantages and disadvantages of different algorithms.In addition, we studied the basic principles of the subspace speech enhancement algorithm and made a simulation experiment for the subspace method based on GEVD (Generalized EigenValue Decomposition). At the same time, we made an analysis of the reasons why the existing speech enhancement algorithms cannot improve the speech intelligibility. Particularly we studied the effects of different kinds of the speech distortions have on the intelligibility in order to design a post filter based on the classification process for the speech distortions. And then combining this post filter with the GEVD subspace method we proposed a signal subspace speech-enhancement method with a high intelligibility performance. The proposed method can constrain the magnitude spectrum of the amplification distortions in excess of6.02dB in order to reduce the damage that these distortions have to the enhanced speech intelligibility. Therefore, the proposed can improve the intelligibility.Finally, we designed a simulation experiment for the proposed method. Moreover we choosed Speech Audiometry (SA) as the intelligibility subjective evaluation method, and selected Normalized Covariance Metric (NCM) as the intelligibility objective evaluation method, and conducted a spectrogram analysis. We used the above three methods to verify the simulation experimental results which show that:the proposed high intelligibility subspace method based on the speech distortions classification can indeed improve the intelligibility of the enhanced speech especially in low SNR conditions.