Study Of Methods Of Speech Feature Extraction

The speech recognition is one of the important research directions of speech signal processing. The study of speech recognition is to force machine to understand what the people say. with the development of the computer, multimedia, and digital signal processing technology, the more wishes is given to the development of speech recognition. In this paper, the speech recognition system on the basis of isolated word is studied. The learning vector quantization and artificial neural network is applied to training of speech feature parameter, and the algorithm of the Dynamic time warping is used to the recognition process. Before the training, the sequential cluster network is used in time warping of the speech signal,the recognition speed is increased as the dimensions of the feature vector and the transient redundant information are decreased. Three algorithms for endpoint detection in speech signal are analyzed. For ensuring the speech recognition correctly, the double-limit endpoint detection algorithm is adopted in recognition system. Several methods of speech feature parameters extraction, such as Liner Prediction Cepstrum Coefficient (LPCC), Mel-Frequency Cepstrum Coefficient (MFCC), and Discrete Wavelet Transform Coefficient (DWTC) are analyzed, and the performances of the recognition algorithms based on these parameters are compared. A new recognition algorithm using both MFCC and DWTC is given, and the validity of this algorithm is verified by simulation.