Study On Objective Quality Comprehensive Evaluation Of Face Mask Speech

Full face mask is widely used in the frog diving,fire rescue,dangerous goods production and many other aspects.The spectrum structure of speech from face mask wearer is different from the speech in normal air.Therefore,the speech quality evaluation algorithm in the air is not necessarily applicable to face mask speech,and speech quality is an important indicator to evaluate the effect of speech enhancement system.Therefore,it is particularly important to study the suitability of quality evaluation algorithm to face mask speech.A comprehensive evaluation of quality and intelligibility of the enhanced face mask speech is made in the thesis,the different effects of face mask speech and air speech under the same conditions are compared,and the suitability of five kinds of objective evaluation algorithms are analyzed,the influence of three speech enhancement algorithms on mask speech intelligibility are compared.The main contents of this thesis are as follows:(1)Face mask speech enhancement.The face mask speech and air speech are added with different signal-to-noise ratio pink noise and wave noise,then enhanced by Wiener filter and logarithmic spectrum mean square error(LSA-MMSE)algorithm respectively.Firstly,the different characteristics of mask speech and air speech are analyzed by spectrogram,then enhancement procedure of two kinds of speech under the same SNR and background noise condition are showed through spectrogram,the enhancement effect is analyzed.(2)Analyze the suitability of objective evaluation of speech quality of face mask speech and air speech.The experiment results of segmental signal to noise ratio(SegSNR),logarithm spectrum distance(LSD),mel cepstrum distance(Mel-cd),modified bark spectrum distortion(MBSD)and perceptual evaluation of speech quality(PESQ)are presented.The experiment results of above five kinds of objective quality evaluation are analyzed in detail.The results show that MBSD is suitable for the evaluation of mask speech enhanced by LSA-MMSE under pink noise,which is not suitable for evaluation of mask noise under wave noise.PESQ is suitable for evaluating mask sounds in wave noise environment but is not suitable for evaluating WF-enhanced air speech in pink noise.SegSNR and LSD are the most widely used,Mel-cd is not suitable for the evaluation of two kinds of speech.(3)An improved algorithm based on ideal binary mask(IdBM)is presented by combining a noise tracking algorithm.By setting the mask threshold from the binary value to the ratio value mask algorithm which changes according to the noise estimation result of each frame,a low complexity noise tracking algorithm is used in the noise power spectrum estimation to enhance the speech,the results show that the weighted function MMSE algorithm can estimate the noise more accurately.(4)Influence of speech enhancement algorithm on mask speech intelligibility.For the mask speech under pink noise and wave noise,Wiener filtering,ideal binary mask and ideal ratio mask are used to improve the mask speech intelligibility.Using the short-time objective intelligibility(STOI)algorithm to evaluate the enhanced speech.Experiment results show that Wiener filtering can only improve the quality of face mask speech,but the speech intelligibility is decreased.The ideal binary masking can greatly improve the intelligibility of face mask speech,and the score of STOI using ideal ratio mask is better than IdBM.