The Research Of Speech Features Extraction Based On Invariant Sets Multi-wavelet

Presently, most of the speech recognition systems perform very well in quiet environment, but in noisy environment the system performance will have a great decline. This issue has become a major obstacle of the practical application of speech recognition system. Therefore, antinoise speech recognition research has gradually become the focus in the field of speech recognition research. Feature extraction is one of the especially important steps in the speech recognition system. Due to the difference of recognition and antinoise performance, different parameters will have a direct impact on the noise robustness of the system.This paper briefly reviews the common speech feature firstly, analyzes the extraction of the current widespread used feature in speech recognition– the Mel-Frequency Cepstrum Coefficient (MFCC) and simulates it. And then we introduce the invariant sets multi-wavelet theory proposed by Charles A.Micchelli and Yuesheng Xu. On this basis, we talk about the derivation of fractal multi-wavelet filter with the interval characteristic of self-affine and the usage of it. According to the ordinary derivation method of biorthogonal multi-wavelets filter, we construct biorthogonal multi-wavelet filter based on the triangle domain, and illustrate the multi-wavelet decomposition can accurately reconstructed without borders distortion effects. Considering the invariant sets multi-wavelet perfectly combine smoothness, short support, orthogonality, symmetry, etc. For short-time fourier transform having inevitable disadvantages in speech feature extraction , as a result of study on multi-wavelet theory and the extraction of MFCC, we use the multi-wavelet transform to replace the short-time fourier transform and Mel filter, and then get a new speech feature MWBC. Also we simulate the new feature MWBC.Finally, the paper introduces the basic theory of speech recognition and hidden markov model, and studys HMM's parameter selecting such as type, structure, states and mixtures. Experiments of Chinese numeral recognition in quiet and the additive Gaussian random noise environment show that the recognition performance and anti-noise performance of the new feature we proposed are better than MFCC's. It provides a new way to improve the noise robustness of speech recognition system.