Robust Features Derived From Differentiated Phase Autocorrelation Spectrum for Speech Recognition

Free registration required

Executive Summary

The performance of the traditional feature extraction methods such as Mel Frequency Cepstral Coefficient (MFCC) degrade in the presences of noise. To overcome this problem many feature extraction methods have been proposed but in most of them the performances of the features degrade in clean condition. One of these methods is Phase AutoCorrelation (PAC) based features that works good in noisy condition but in clean condition the recognition performance is low. In this paper windowed autocorrelation coefficients and differentiated PAC spectrum have been used to overcome this drawback.

  • Format: PDF
  • Size: 175.04 KB