Research And Improvement Of Anti-noise Method For Speech Recognition Based On HMM

With the development of technology and the wide application of computer, people want to be able to communicate with the computer in a more direct and rapid way, which can understand the natural language of mankind, speech recognition technology was born and has made considerable progress in order to achieve this aim. At present, the speech recognition has been quite mature and achieved good results in no noise environment. Speech recognition encounters a few bottle-neck that the effect of recognition will be greatly decreased considering the noise environment, the matching degree of recognition template and training template are not high and lower recognition rate because of the complexity of noise and its influence on the speech parameters. Therefore, we should remove the biggest obstacle of environmental noise, and further research anti-noise robust speech recognition system, in order to maximize the practice of speech recognition.Aiming at the problem of environmental noise, robust speech recognition technology was studied. With the research and development of robust speech recognition technology, the existing major robust speech recognition are speech enhancement, robust speech feature extraction and robust speech model parameters adjustment. In order to make the speech recognition system to achieve better recognition performance, researches of speech recognition focus on using reasonable technology and can combine the above technology effectively based on different environmental noise. Research is focused on robust speech recognition problem of the speech enhancement stage and the speech recognition feature extraction stage in this thesis.The basic principle of speech recognition system was introduced first. Speech recognition system consists of preprocessing, pattern recognition and matching, feature extraction. Because the wavelet transform has the local signal good analysis ability not only in time domain but also in the frequency domain. So it is a better method of signal analysis, and is widely used in many fields such as signal de-noising. The analysis method of bionic wavelet on the basis of wavelet transform was introduced and the bionic wavelet transform theory was studied adequately, combining the characteristic of bionic wavelet coefficient correlation, research on bionic wavelet correlation de-noising method.From the point of view of practical application, we put forward a de-noising method based on bionic wavelet transform and correlation algorithm on basis of analysis and research about traditional speech enhancement method, the simulation of Matlab results show that the method has better performance combining with their advantages and disadvantages. In the speech feature extraction stage, a new robust noise characteristics parameter BWTMFCC based on bionic wavelet transform was proposed for speech recognition.A simple non specific person, small vocabulary speech recognition system was built by using software platform. The speech enhancement algorithm and the new robust noise characteristics parameter BWTMFCC based on bionic wavelet transform were used in the speech recognition system. Compared with the recognition rates of different characteristic parameter of the system through experiment, the effectiveness of the algorithm was verified finally.