Robustness analysis of automatic speech signal recognition system against factors degrading speech signal

Abstract

In the article there are presented the results of research on the influence of the lossy compression, used in codecs G.711, G.723.1 and iLBC, on the efficiency of isolated speech phrase recognition. In the research the degree of robustness against degrading factors in the parameterisation method of audio signal LPCC and MFCC (Linear Prediction Cepstral Coefficients, Mel Frequency Cepstral Coefficients) is compared. The research is based on the classifier of improved Gaussian mixtures making allowance for Universal Background Model GMM-UBM (Gaussian Mixtures Model - Universal Background Model). The research was conducted on the database composed of 3000 isolated speech phrases.