An experimental comparison of different feature extraction and classification methods for telephone speech

Robust speech recognition over telephone lines severely depends on the choice of the feature extraction and classification methods. In order to get the highest possible performance of the speech recognizer a number of commonly used feature extraction methods (MFCC, LPC, PLP, RASTA-PLP) and classification methods (MLP, LVQ, HMM) were tested on the same telephone speech data. All combinations of feature extraction and classification methods were computed and several parameters of both methods where changed in order to find a non-local maximum of recognition accuracy. The paper does not describe a comparison of classification but of feature extraction methods because it is clear that an HMM would outperform both LVQ and MLP. The big question is if the same feature extraction methods always lead to the best results, no matter which classifier is used!