Research And Implementation Of Dual-microphone Speech Enhancement Algorithm Based On Power Level Difference

In practical application environments, speech signals may be affected by many kinds of noises, these noises will seriously damage the speech quality. The purpose of speech enhancement is to eliminate the interference of noises and other interference sources to the target speech, extract the pure target speech, improve the speech quality, and increase the intelligibility and comfort of speech.According to the number of microphones, the existing speech enhancement algorithms can be divided into three categories, i.e. the single channel speech enhancement algorithm, the microphone array speech enhancement algorithm and the dual-channel speech enhancement algorithm. In general, the speech enhancement algorithms based on small microphone arrays are more practical with consideration of the system complexity and cost. In the smart phones, the speech enhancement algorithms based on small microphone arrays are most commonly applied at present.In this thesis, the dual-microphone speech enhancement algorithm for mobile applications has been mainly studied.Firstly, the research background, significance and research status of speech enhancement are summarized, the single channel speech enhancement algorithm, the microphone array speech enhancement algorithm and the dual-channel speech enhancement algorithm are studied systematically. Besides, the classical algorithms of these three classification are introduced and evaluated simply.Secondly, a dual-microphone speech enhancement algorithm based on energy difference is studied in detail. The algorithm is not sensitive to the time delay estimation and has good performance, thus it is quite suitable for the smart phone application. In the traditional power level difference(PLD) algorithm, the fixed smoothing factor applied in signal spectrum estimation, and the stationary noise estimation methods uesd. Thersfore, its performance may become very poor in the higher non-stationary noise environment. In addition, this algorithm will introduce musical noise in the processed speech when SNR is lower, this will affect hearing comfort of the speech. Based on the existing problems within the traditional PLD algorithm, an improved PLD algorithm is proposed in this thesis. In the improved PLD algorithm, an adaptive smoothing factor of signal spectrum is applied to increase the accuracy of signal spectrum estimation, a noise estimation method based on standard energy difference is used to enhance the accuracy of noise estimation in non-stationary noise environment, and a post filter is combined to deal with the musical noise. In this thesis, the traditional PLD algorithm and the improved PLD algorithm is implemented with MATLAB programming language, based on the ideas of the two algorithms.Finally, performance of the improved PLD algorithm has been tested and analyzed. In order to do that, one single channel spectral subtraction algorithm, the traditional PLD algorithm and the improved PLD algorithm are used to process the same speech signals degraded by background noises. Then the process results include perceptual evaluation of speech quality(PESQ) values and spectrograms are compared and analyzed. The results have shown that the dual-microphone speech enhancement algorithm has obvious advantages over the single channel speech enhancement algorithm. Compared with traditional PLD algorithm, the improved PLD algorithm gains the higher PESQ values, improves the musical noise phenomenon, and makes speech being heard much more comfortable than the two others. However, there are still some deficiencies in the improved algorithm, which will be mentioned in the conclusion part of this thesis.