| In the information society, digitized voice transmission, control and identification have become one of the basic components of digital communication sections. However, the voice signal would be inevitably disturbed by various types of noises in the process of acquisition and transmission. Not only would the quality of received voice be degraded, but also the normal works of the speech recognition and control system would be affected. Along with the fact that voice digital signal processing technology has come into wide practical use, speech enhancement technology has become one of the most important problems which should be urgently solved in this period. Speech enhancement is designed to eliminate noise pollution and improve the speech intelligibility. For different types of noise interference, different speech enhancement methods should be used to suppress the background noise and improve the listeners’comfort as well.Based on excellent research results of many scholars in the field of speech enhancement, the research contents of this paper follow a progressive relationship, which are summarized as follows roughly:1. The basic principles and common methods for speech enhancement are briefly described in this paper.The nature of various noises and their interferences on the voice are analyzed.2. For the case of non-stationary noise, secondary smoothing is introduced into the Voice Activity Detection (VAD) algorithm in post-processing, which aims to solve the bias caused by VAD while stationary noise is estimated. Wiener filtering is used for estimating pure voice instead of spectral subtraction, which avoids generating the "musical noise". This algorithm can accurately estimate the noise and get a better enhancement effect while balancing the complexity and treatment effect. The applicability of improved algorithm is analyzed by adding a variety of non-stationary noises, the results show that the algorithm is more suitable for processing stationary noise.3. As for the complex situation of non-stationary noise, the Data Driven Recursive (DDR) method is analyzed in this paper. Vuvuzela, babble, train and car noises are respectively used to simulate the algorithm. The treatment effectiveness of noise pollution by this algorithm is proved. It also proves that the enhanced algorithm proposed by this paper has a good balance between the complexity and effectiveness. The law has been found out that some algorithms which are suitable for stationary noise maybe would not work well under the condition of non-stationary noise, but algorithms which are suitable for non-stationary noise certainly can work out perfectly in the case of stationary noise. The effective achievement of DDR algorithm provides support to the Ideal Binary Mask (IBM) algorithm in the later part of this paper.4. Improving the intelligibility is one of the major purposes for speech enhancement. The IBM algorithm and Harmonic Retrieval(HR) algorithm which can improve the intelligibility are analyzed in this paper. The IBM algorithm is realized based on the DDR method which is used to estimate the noise variance. The effectiveness of this algorithm is proved by simulated results. In this paper, the HR algorithm is improved by3class sub-bands processing in frequency domain, which solves the problem of spectrum aliasing caused by the convolution when traditional HR method is used. The output signal which is enhanced by IBM method is used as the input signal of the improved HR method for the twice enhancement processing, which can improve the speech intelligibility effectively. |
PDF Full Text Request