Research And Implementation Of Speech Enhancement Based On MMSE Amplitude Spectrum Estimator

Speech enhancement is important in the field of speech signal processing,which is aimed to improve the quality and intelligibility of speech signal in low signal to noise ratio(SNR)condition before further processing of speech.In low SNR or highly unstable noise environment,the quality and intelligibility of speech will decrease quickly,thus human or machine cannot recognize or transmit voice very well.So it is important to improve the quality of speech signals for speech processing,and speech enhancement is of great significance and practical value.In this thesis,an analysis is given on state-of-the-art speech enhancement algorithms,then the speech enhancement algorithm based on minimum mean square error(MMSE)is studied,and measures are taken to improve its performance;thereafter,MATLAB toolbox is utilized to evaluate the performance of the improved algorithm.Finally,the improved algorithm is implemented on the DM642 device to build up an audio acquisition,processing and playing system,and fixed-point optimization is performed on the improved algorithm to elevate its real-time performance.The main content of the thesis is as follows:(1)The advantages and disadvantages of all kinds of current speech enhancement algorithms are analyzed.And the speech enhancement algorithm based on MMSE amplitude spectrum estimation is chosen as research objective,whose disadvantages such as deficiency in a prior SNR estimation,large noise estimation error,as well as poor real-time performance of DSP implementation,etc.,are pointed out.(2)To tackle with the delay of Decision-Directed(DD)algorithm,a new algorithm named DD Based on Single frequency Entropy(DDBSE)is proposed in this thesis for a prior SNR estimation.The algorithm is based on the change of entropy,which is able to track the change of speech quickly,thus improving the performance of DD algorithm and providing more accurate estimator for the amplitude of SNR estimation.(3)Aimed at the insufficient in noise estimation,a new noise estimation algorithm called UMVAD is presented,in which the unbiased MMSE and improved statistical voice activity detection(VAD)are combined to minimize under-estimation and over-estimation of noise.The experimental results prove that this algorithm can provide minimum value estimation of median square error.Speech enhancement based on DDBSE and UMVAD can obtain better performance on segSNR and composite score,especially in low-SNR and non-stationary noise environment.(4)A double thresholds VAD decision method is put forward to improve the VAD algorithm used in the noise estimation.Compared with several typical phonetic characteristics,the improved statistical VAD has combined the characteristics of energy-entropy(EE)and mean-Delta(MD),can achieve more satisfactory overall performance.At the same time,an environmental awareness speech enhancement algorithm is proposed based on the idea of noise classification by introducing noise classifier and support vector machine(SVM)to optimize the adjustable parameters in the algorithm.(5)The proposed speech enhancement algorithm has been implemented on DM642 DSP chip to build up an audio acquisition,processing and playing system,and fixed-point optimization is performed on the improved algorithm to elevate its real-time performance.Compared with pristine floating-point algorithm,the processing speed of the optimized fixed-point algorithm has been promoted about 83%.And the audio acquisition,processing and playing system can process the speech signal in real-time,in which a frame of speech can be handled at about 8 milliseconds.