Research On The Anti-noise Methods In Speech Recognition

Currently, speech recognition systems have obtained great achievements in abetter environment. However, the interference signals exist in practical application,making the performance of speech recognition system decrease significantly. Thus,de-noising technology has become a key of whether the speech recognition systemcan perfect application in life, and it is also a hot issue to be addressed in the field ofspeech recognition. Currently, the main anti-noise technologies, including speechenhancement, anti-noise feature extraction and model compensation. This paperpresents a combination of de-noising method, which combines two technologies ofspeech enhancement and anti-noise feature extraction to improve the robustness of thesystem.Firstly, research on speech enhancement technology, by analyzing the advantagesand disadvantages of hard threshold function, soft threshold function, soft and hardthreshold compromise function and Garrote threshold function, to construct animproved threshold function. This function has the advantages of the above functions.Then verify the feasibility and effectiveness of the function through simulation.Secondly, we usually use MFCC parameters and improved MFCC parametersbased on wavelet multi-resolution analysis, when extraction anti-noise featureparameters. The analysis window of FFT transform isn’t change in time domain andfrequency using MFCC parameter, this is a violation of the non-stationary nature ofthe speech signal; while wavelet multi-resolution analysis only disconnect the lowfrequency part. This paper proposes a new method base on the wavelet packetanalysis for these two defects, and it has a better recognition rate.Finally, we construct a speech recognition system based on combinationde-noising method of this paper. We have demonstrated the effectiveness of this newmethod, through comparison the identify rate of different systems in several differentSNR environment.