关键词: 语音信号处理, 稀疏分解, 匹配追踪, 语音识别

Abstract: After studying Matching Pursuit（MP） based speech signal sparse decomposition，a new sparse decomposition algorithm is proposed.According to the characters of the speech signal and the fact that the signal and atoms are all real，based on the partition of the over-complete atom dictionary，this new algorithm converts inner product calculations into crosscorrelation calculations that are fast done by Fast Hartley Transform（FHT）.Therefore，compared with MP based signal sparse decomposition with FFT，this algorithm can not only reduce the memory consumption by half，but also heighten the speed of the decomposition by 24.8%.The speech recognition is realized by Support Vector Machine（SVM） that uses speech signals’ Mel-scaled Frequency Cepstrum Coefficient（MFCC） and the features extracted by this new algorithm.Finally the experiments verify the effectiveness of the proposed algorithm.

Key words: speech signal processing, sparse decomposition, Matching Pursuit（MP）, speech recognition