Empirical Mode Decomposition Method And Its Research In Speech Recognition Algorithm

With the rapid development of the discipline, such as physiology, acoustics, electronictechnique, pattern recognition, signal processing and so on, the field of speech recognition ispushed to a whole new level. Speech recognition technology has shifted from laboratory topractical application now. But the speech recognition rate could be still improved in thepractical application. The accuracy from theory algorithm should be improved efficiently. Thetraditional methods of time-frequency analysis are based on Fourier transform, such as shorttime Fourier transform, wavelet transform and so on. Fourier transform is often used instationary and linear signal, but the speech signal is non-stationary and non-linear signal. Sothese methods can not solve the disadvantages of Fourier transform fundamentally. Therefore,empirical mode decomposition algorithm which used instantaneous frequency as basic unit isresearched in this paper. Under different signal-to-noise ratio, detecting the speech endpoints isthe main purpose of this paper. The further purpose is to identify the speech signal. It has apractical significance for the application of speech.In this paper, key points of main work are as follows:1. The basic concept of empirical mode decomposition algorithm which proposed byHuang N E and its decomposition are studied. While EMD algorithm is applied in speech signal,a set of intrinsic mode functions and a residual are received after the speech signal decomposed.It’s illustrated that the EMD algorithm is adaptive, multi-resolution and local features.2. Speech endpoint detection is the front of speech recognition system. It is used fordetecting the effective speech signal start and end point under the noise environment. Theresults will affect the speech recognition rate. They are combined with average magnitudedifference function and mel-frequency cepstral coefficients respectively. Then, two novelspeech endpoint detection based on EMD algorithm are proposed: EMD+AMDF andEMD+MFCC. Computer simulations analyze the speech endpoint detection performance from EMD+AMDF and EMD+MFCC. Compared with the endpoint detection results from traditionaldouble threshold method, it proved that the proposed algorithms are effective. It’s achievedspeech endpoint detection in low SNR environment.3. According to the speech recognition problem and the noise influence, a speechrecognition algorithm using EMD is designed. The method uses EMD algorithm to decomposespeech signal. After that, the MFCC of intrinsic mode functions are extracted as speech featureparameters. The identification model is RBF neural network model. The speech recognition rateis improved effectively, and the influence of noise is reduced. The theoretical simulationverifies the effectiveness and accuracy of the algorithm.