Research On Improved Speech Recognition System In Low SNR Environment

Speech recognition in low SNR environment is an important research area in the world.It is also a focus issue and difficulty in speech recognition field,which has important theoretical and practical significance.Based on the research of some basic theories of speech signal processing,the thesis focuses on the methods of speech signal noise reduction and enhancement,speech signal start and stop endpoint detection,feature extraction and speech recognition.Based on the study and research of speech recognition system,the thesis studies the noise immunity of the existing main algorithms,and discusses the new anti noise speech recognition algorithm.For the traditional spectral subtraction in the low SNR environment,it is easy to distort the speech waveform and introduce new noise.The principle of empirical mode decomposition(EMD)and multi-window spectral estimation algorithms is analyzed in-depth,and combined with their respective advantages of the EMD is proposed based on the improved multi-window spectrum estimation spectral subtraction speech enhancement algorithms.Firstly,the EMD method to decompose the noisy speech into different frequency bands is used on the algorithm.Then,multi-window spectral de-noising on the noise-free intrinsic mode function(IMF)in different frequency bands is performed.Finally,EMD reconstruction on the processed IMF to obtain Speech enhanced signal is performed.Experimental results show that this method can effectively eliminate voice background noise and "music" noise,and maintain high voice quality.On the basis of analyzing the feature of speech endpoint,the end point detection method based on cepstrum distance is studied,and a new endpoint detection algorithm based on adaptive dynamic threshold threshold method based on LPCC and MFCC cepstrum distance is studied in the thesis.Firstly,the algorithm analyzes the advantages of the LPCC cepstrum distance endpoint detection method and the feasibility and scientific basis of the MFCC cepstrum distance method applied to the endpoint detection,and then compares the pros and cons of the respective algorithms,and proposes a cepstrum parameter extraction algorithm combining MFCC and LPCC methods.Finally,an improved adaptive dynamic threshold method is used to improve the endpoint detection algorithm.Experimental results show that the detection accuracy of the vocal endpoints can reach over 98%.The proposed algorithm is also very effective in combination with this issue's noise reduction algorithm in the low SNR environment.The basic principle of Hidden Markov Model is analyzed in theory and its three important algorithms: forward-backward(FB)algorithm,Baum-Welch algorithm,and Viterbi algorithm,and proposes an effective solution for the emergence of effective State number and data overflow problems.A speech recognition system based on the HMM model is built on the MATLAB platform,and the effects of the improved algorithm on the performance of the recognition system are compared and analyzed in detail.