Study Of Noise-robust Methods For Automatic Speech Recognition

Automatic speech recognition system robustness in noisy environment is the key to speechrecognition system from laboratory to practical application, the main objective of the study is toidentify the performance between application environment and training environment mismatchcaused by falling. Based on the summarization and analysis of existing noise robust speechrecognition methods, this dissertation studies the algorithms in signal space and feature space,including the endpoint detection, speech enhancement and robust feature extraction. The mianresearch work and innovations are listed as follows:Firstly, this thesis introduces the digital model of speech signal and the basic principle of speechrecognition, then according to the composition of the speech recognition system, we summarizedand classified the noise robust methods in automatic speech recognition into the signal space,feature space and model space.Then, endpoint detection algorithms such as short-time energy, short-time zero-crossing rate andspectral entropy are studied. On the basis of such algorithms, a new endpoint detection algorithmbased on improved spectral entropy is proposed. The simulation results shows that the proposedalgorithm is superior to the basic spectral entropy algorithm and has better noise robustness.Furthermore in order to suppress the noise effect on speech signal, the spectral subtraction andimproved spectral subtraction are researched.Secondly, this paper studies feature extraction in robust speech recognition from the featurespace, and analysis three kinds of feature extraction parameters commonly used in speechrecognition system, LPC, LPCC and MFCC. This paper proposes a robust feature parameters basedon MFCC: SS-MFCC. The recognition experiment can get new characteristic parameters withrespect to the improvement of the degree of MFCC in different noisy environment, which verifiesthe good robustness of new parameters.Finally, using Matlab, thesis establishs a speech recognition system based on the HiddenMarkov Model. Experiments show that, for Chinese digit0-9, the recognition rate of the speechrecognition system which we build with above algorithms can reach83.75%in condition of SNR is10dB. And the system has good robustness.