Discriminative Output Coding Features for Speech Recognition

This paper presents a novel approach of discriminative acoustic feature extraction for speech recognition using output coding technique. A high dimensional feature space for higher discriminative capability is constructed by expanding MFCC coefficients with polynomial expansion. In order to fit the discriminative features in the hidden Markov model structure of speech recognition, the high dimensional feature vectors are further projected into a low dimensional feature space using the output scores of a set of SVMs. Each of the SVMs is trained in one phone versus the rest manner so that each of the resulting feature dimensions can provide effective information to differ one phone from the others. The discriminative features have been evaluated in the speech recognition task of the TIMIT corpus, and 72.18% phone accuracy has been achieved.