The Research For Dimensional Reduction And Recognition Of Chinese Digit Based On Speech Feature

Chinese digit speech recognition is to identify the speech of speaker-independent Chinese digit from"0"to"9". Feature extraction of speech signal is the important precondition and foundation of speech recognition. This paper focuses on the feature extraction processing, and the main research works are as follow:First, GMM clustering is used to reduce the data of MFCC for Chinese digit identification. MFCC is widely used in Chinese digital identification. Because the amount of MFCC extracted from 0-9 is too large, the mean of model parameters which is clustered with GMM by MFCC to reduce the amount is employed as a new feature with DTW for Chinese digital identification. Simulation results demonstrate that the amount of the new feature is 30.9% to that of MFCC, the running time reduces by 207.12s, but the recognition rate decreases by 1.11%.Second, Locally linear embedding algorithm is used to extract the feature of speech signal for mandarin digit speech recognition. Speech signal dimensions are higher when the signal is transformed to frequency domain, manifold learning algorithm can find a smooth low-dimensional manifold embedded in the high-dimensional data space. In this paper, manifold learning algorithm is proposed to reduce the dimensions in the high-dimensional data for mandarin digit speech recognition. First, low-dimensional manifold structure is extracted from the high-dimensional frequency data based on locally linear embedding of manifold learning algorithms. Then the resulting low-dimensional data is inputted into DTW to recognize. Simulation results demonstrate that the dimensions are lower using LLE compared with MFCC, the recognition rate increases by 1.2% in mandarin digit speech recognition, and the recognition speed gets improved effectively.Third, genetic algorithm is used to reduce the feature dimension for mandarin digit recognition. The dimensions are higher after combining MFCC with LPCC. In this paper, genetic algorithm is proposed to reduce the dimensions of the feature data to improve recognition performance of the system. First, extract MFCC and LPCC of the speech signal; then, reduce the dimensions of the feature data based on genetic algorithm; finally, the low dimensional data are sent into the support vector machine. Simulation results demonstrate that the recognition rate increases by 12.2% using genetic algorithm compared with principle component analysis, recognition rate almost has no change compared with the initial characteristics and the recognition speed gets improved effectively.