Accuracy of Feature Extraction Approaches in the Task of Recognition and Classification of Isolated Words in Speech

The paper presents the practical comparing of feature extraction methods in speech by the example of isolated words (commands) recognition and comparing the performance of a set of features using different classification algorithms. In this case, a consistently increasing set of keyword samples is used for the training set. Approaches based on cepstral coefficient extraction (including MFCC- and GFCC-based variations), linear prediction (LPCC), and wavelet transform (via Basilar-membrane Frequency-band) are considered. The obtained results indicate that in some cases (including recognition tasks) the use of features other than MFCC allows to increase the accuracy of recognition from 1 to 7 percent. At the same time, the statements of the authors of other approaches to the extraction of speech features about the significant superiority of their methods over MFCC are not confirmed.