Study On Speech Enhancement Algorithm Based On Minimum Statistics

Speech is an important information carrier when human beings communicate with each other. But lots of speech communication systems are unavoidably influenced by different kinds of noise such as road noise, fan noise and noise from some high power equipments. Severe noise interference will reduce the clarity and intelligibility of the speech and deteriorate the communication quality.Speech enhancement is one of the major techniques of speech processing. It’s a commonly used method to attenuate the noise interference and its purpose is to extract the pure speech from the noisy speech signal and increase the signal to noise ratio (SNR). Speech enhancement based on short-time spectral estimation is a widely-used frequency domain algorithm of speech enhancement. How to estimate the background noise spectrum is one of the crucial components of this method. And most of the current algorithms cannot have satisfactory performance when dealing with non-stationary noise. The main purpose of this paper is to investigate the cost-effective way to improve the tracking capability of the noise spectrum estimation method.The noise spectrum estimation result is directly related to the performance of the speech enhancement system. We focus on the noise spectrum estimation algorithm based on the minimum statistics which is widely-used in speech enhancement. This algorithm has clear physical foundation, moreover, it can estimate the background noise whether the voice is active or not. The problems of this algorithm are:(1) It usually has a two to three second time-delay, which prevents it from tracing non-stationary noise efficiently.(2) Its computation is quite complicated, and in practical application, a constant number is often used as a compensation factor, which is detrimental to the accuracy of the estimation and also harmful to the tracking capability. In order to solve these two problems, we propose a cost-effective and reliable noise segment detection method. This method uses the signal stationarity as the characteristic parameter and set a threshold value which is favorable toward detecting the non-voice segment. With this method we can calculate the newest background noise spectrum which can dynamically compensate the noise estimation result achieved from searching the minimum values of the spectrum. This method can make the noise estimation more accurate and efficiently track the change of the noise, and finally promote the performance of the speech enhancement system.In the last part of this paper, we combine the unidirectional microphone array algorithm with the improved noise estimation algorithm. The problem of malfunctioning when the level of background noise is too high is well solved by the proposed method.