Signal Processing of Noisy Short Utterance Based on Noise Separation and Multiple Features Fusion

Recognition rate of noisy short utterance is lower, the two main factors are the inadequate training data and utterance polluted by noisy seriously. In this paper, we proposed corresponding algorithms. First, noise and speech are regarded as parallel information, we use FastICA algorithm to separate pure speech and noise. And then, we use differences detecting and eliminating algorithm (DDAEA) proposed in the paper to purify speech signal further, so that we can get more pure speech. In the phase of feature extraction, we combine vocal sourceand vocal tract-related characteristics to make full use of limited information, so we get three score evaluation subsystems: MFCC D LPCC+WOWOR4, MFCC D LPCC+WOWOR6, MFCC D LPCC+WOWOR8, the highest score of the three subsystems is the final score of the system. Experiments prove that the above algorithms improve speaker recognition performance of noisy short utterance.