Study On Low-SNR Speech Enhancement And Separation In Driving Environment

With the development of automotive industry,many intelligent modules have been implemented in vehicles,such as automatic driving assistant,multimedia entertainments facilities,head up display,and so on.Voice recognition is one of important technologies to improve the quality of experience for human-machine interactions.However,the noise is the key factor related to the voice recognition performance,and there are complex noises in driving environments,including passenger voice,engine noise,tire noise and wind noise.It is one of very important problems to enhance the SNR of driver's voice signal,to get much precise voice recognition in driving scenario.The characteristics of noises in a low-SNR driving scenario are analyzed in this thesis,which could be classified into two types,as non-voice noise and voice noise.The non-voice noise can be eliminated by some enhancement methods.Voice noise is separated from target driver's speech signal by blind source separation algorithm.In this thesis,the voice enhancement and voice blind source separation algorithms are studied,including the following three aspects:1.In driving scenario,the speech signal has much higher energy in high frequency than the non-voice noise,as shown in analysis.A suppression preprocessing method is proposed to make noise cancellation in low frequency domain.It gets better SNR and PESQ performances,as verified with experiments.2.Based on spectral subtraction method,a smoothing factor optimization algorithm is proposed.Speech presence probability(SPP)is introduced into the calculation of the smoothing factor to estimate the noise adaptively.Furthermore,adaptive first-order recursive smoothing noise estimation algorithm with bias compensation is proposed.The proposed method has a better noise elimination effect than MS,MCRA and IMCRA algorithms in a low-SNR driving environment.3.In blind source separation,the negative entropy-based FastICA separation algorithm contains Newton iteration procedure.But the initial values of the Newton iteration will impact the performance of the separation algorithm.The Newton downhill and speech separation algorithms are combined,in which the downhill factor is self-adaptive.The iterative time and number of iterations of the improved algorithm have been reduced.The proposed algorithm reduces the sensitivity of Newton iteration to initial value.As verified by experiments,the proposed algorithms have improved the target driver's speech quality,and much suitable to the speech enhancement and separation in the low-SNR driving environment.