Research On The Application Of HHT Time-frequency Analysis In The Overall Recognition Of Chinese Speech

The commonly used time-frequency analysis method based on the Fourier transform theory inherits the deficiencies of the Fourier analysis method in dealing with non-stationary nonlinear signals.Radon transform and wavelet transform,for example,are only suitable for the processing of FM signals and self-similar signals;Wigner-Ville distributions are always affected by cross terms;STFT and Gabor transforms are suitable for signals with short-term stationary characteristics.In addition,these traditional time-frequency analysis methods have poor adaptability.According to the Heisenberg uncertainty principle,their time domain resolution and frequency domain resolution often cannot reach a high degree at the same time.The HHT method proposed by Huang et al.has been widely used due to its good self-adaptability and ability to accurately describe the variation of non-stationary signal frequency with time.The HHT time-frequency analysis method includes two steps of Empirical Mode Decomposition(EMD)and Hilbert Spectrum Analysis(HSA).First,the EMD process adaptively decomposes the signal into a series of finite number of Intrinsic Mode Functions(IMF),which makes the concept of instantaneous frequency used in the following HSA have practical physical meaning.By HSA,the energy of the signal is expressed in the time-frequency joint plane,on which each specific moment and instantaneous frequency has a certain signal function amplitude,and the accuracy is higher than that of the Fourier spectral map.The speech signal is a typical non-stationary signal.Combined with the advantages of the HHT method,this paper proposes a feature extraction algorithm based on the HHT method for feature extraction and classification recognition of Chinese speech signals.First,the Chinese speech signal is processed using the EMD algorithm.Then the first two IMF components are selected to obtain the Hilbert spectral map of the Chinese speech signal.We obtain the Hilbert spectral map of each Chinese speech signal as a visual representation of their respective speech signals.Feature vectors are extracted from the image matrix of the Hilbert spectral map.Finally,statistical analysis and variance test are performed on the recognition ability of the acquired feature vectors,and C-SVC two-class support vector machine model is used to classify the Chinese speech signals.The experimental results show that the proposed method is effective and has a good stability in a small vocabulary: the recognition rate of 24 Chinese phonetic finals can reach more than 90%;the recognition rate of 10 syllable Chinese vocabulary can reach more than 98%;the recognition rate of 40 two-syllable Chinese vocabulary can reach more than 94%.In addition,the experiments show that the method given in this paper can still maintain a high recognition rate and have a certain anti-noise ability under the condition that the speech signal is superimposed on noise.