Research On Speech Feature Extraction And Recognition Based On Improved EMD

Among many speech signal feature extraction and recognition technologies,Meir frequency cepstrum coefficient extraction method has been widely used in various speech signal recognition systems because of its high accuracy,low error and good reducibility.However,due to the constraints of sensor performance,external noise,analog and electric conversion distortion in the transmission process of speech signals,there are a lot of original signals doped with invalid speech signals.Traditional methods to process acoustic signals in these environments will filter out a large number of original signals,and there are some problems in both accuracy and recognition rate.Therefore,this paper studies the feature extraction and recognition algorithm of voice signal.Firstly,this paper theoretically deduces the introduction of speech signal feature extraction,analyzes the physiological model and mathematical model of speech signal generation,studies the process of speech signal preprocessing,and elaborates on the reasons for selecting Hamming window to divide and add Windows to the signal and selecting the double-threshold endpoint detection method.Secondly,the most commonly used Meier cepstrum parameters in the process of feature extraction are theoretically derived,and the empirical mode decomposition algorithm is introduced.Through simulation analysis and comparison,the difference between the invalid noise signal and the useful original signal autocorrelation function waveform is obtained.Based on improved empirical mode decomposition of MEL cepstrum parameter extraction method,by improving the algorithm to distinguish the main modal component signal types,makes finally extract the feature parameters of retain more original signal,this paper operation process on the nonlinear characteristics of speech signal and high compatibility,solve the current weak resistance to noise of the traditional feature extraction algorithm,filtering the original signal,reducing some problems.Finally,in order to solve the problems of low recognition rate caused by inaccurate excitation function of traditional speech signal recognition training network or poor structure of supervised learning network,a speech recognition method based on improved hidden Markov structure is proposed.This method combines the feature that the training value of Hidden Markov pair has nothing to do with the state of the previous moment and the advantages of the mixed Gaussian model.By observing the probability distribution of each moment of the signal,and then optimizing the transfer probability of the training network,the time sequence of the speech is modeled to complete the construction of the speech signal recognition system.Simulation and experimental data processing and analysis verify the effectiveness and feasibility of the proposed scheme.