Research On Speech Enhancement Algorithms In Low Signal-to-Noise Ratio Scene

As an important medium for people to communicate and express their emotions,speech is always disturbed by noise in daily life,so we need to enhance the speech with background noise.The ultimate goal of speech enhancement algorithm is to suppress the background noise,improve the quality of speech auditory,while ensuring a certain degree of speech intelligibility.The speech enhancement algorithm has been studied for more than half a century.During this period,many classical speech enhancement algorithms,such as spectral subtraction,wiener filtering method and minimum mean square error of amplitude spectrum algorithm have emerged,which have been studied so far.These algorithms usually can achieve good speech enhancement performance in high signal to noise(SNR)and stationary noise environment,but the speech enhancement effect is unsatisfactory in low SNR and non-stationary noise environment,so there are many problems that need to overcome.Therefore,the speech enhancement of the noisy speech signal in low SNR and non-stationary noise scene is still a hotspot at home and abroad for scholars.In this thesis,we focus on the improvement of LSA-MMSE algorithm and signal subspace algorithm in low SNR scene.The main research work is as follows:Firstly,an improved LSA-MMSE algorithm in low SNR scene is proposed.For the traditional LSA-MMSE algorithm in strong noisy environment,the effect of integrity about speech.information is poor.The theory that Loizou et al proposed,which most of the speech enhancement algorithm generally existed two different types of distortion after enhanced is applied to the LSA-MMSE algorithm.We made some improvement about LSA-MMSE algorithm based on this theory.Formerly,scholars always classified the attenuation distortion of region ? and the amplification distortion of region ? less than or equal to 6.02dB corresponding to the amplitude spectrum as a class to process,which is considered not to affect the integrity of speech information,but the study found that it would instead result in more residual noise.Based on this point,the attenuation distortion,less than or equal to 6.02dB amplification distortion and greater than 6.02dB amplification distortion corresponding to the amplitude spectrum were taken to varying degrees of downward constraints.In addition,the estimation error of prior SNR and gain function have a great effect on the speech enhancement in low SNR scene,so the improved LSA-MMSE algorithm adjusts them respectively.The experimental results show that the proposed algorithm can better preserve the main information of the speech and effectively suppress the background noise of low frequency part in low SNR scene.Secondly,.an improved signal subspace speech enhancement algorithm is proposed in low SNR scene.Though the subspace algorithm has a good denoising effect,it still remains a lot noise in low SNR environment.In this thesis,we Firstly apply the method that filtering the eigenvalues less than zero and the corresponding eigenvector to traditional subspace algorithm,in order to achieve the effect of optimizing signal subspace.At the same time,the covariance estimation method using shared sine spectrum is proposed to reduce the estimation error and computational complexity.Finally,wiener filter function is introduced to modify the estimated clean speech.The experimental results show that the improved algorithm can effectively remove background noise and improve speech quality in low SNR scenarios of five common noises,and its speech enhancement effect is better than improved before.