The Research Of Robust Speech Recognition In Noise Environment Based On MFCC

Robustness for speech recognition system in noisy environment is the key of speech recognition utility. It is a hot spot and difficulty of the speech recognition research field. The root of the noise robustness problem can be attributed to the speech recognition training and testing environment does not match. Noise robust speech recognition is designed to eliminate noise caused by the training environment and test environment does not match. There are four kinds of methods: robust feature extraction, speech enhancement, model compensation, microphone array. This paper focuses on the robust feature extraction method. It use the auditory characteristics mechanism and the existing window and sub-band technology. Base on this, this paper try to build a strong robust feature vector to enhance the recognition performance in noisy environment, enable it to good use in the actual environment. This paper mainly based on the MFCC(Mel Frequency Cepstrum Coeficients,MFCC) feature extraction, improve the normal MFCC feature extraction process.The main tast of this parper are described in details as follows:Firstly improve the window function,add the mixed-window function in MFCC extraction process instead of the Hamming window. In confirm the width of lobe has not great changes in circumstances, adopt a function that more sidelobe suppression. Such a window function have more advantage than the traditional Hamming window in the noise environment.Then, Apply the sub-band frequency spectrum centriod(sub-band frequency spectrum centriod,SSC) in feature extraction. This process is based on the traditional MFCC feature extraction, add the information of the spectrum peak position which interfered by noise smaller. We calculated the spectrum center of mass which polluted by noise a lesser for very sub-band, according to calculated center of mass spectrum sequence to get the new parameters. The main task of this stage is to determine how many sub-bands divided by the whole band. as well as various sub-band to confirm the location of the border.Finally, combined of window and sub-band frequency spectrum centroid, By applied the sub-band center of mass which describe the spectum peak position information to the extraction process, set up the recognition system based on HMM model. In this paper, we use the HTK speech recognition tools provided by the University of Cambridge, United Kingdom for simulation. Experimental results show that the the recognition performance and robustness of improved feature extraction algorithm has some improvement compared with the benchmark system.