A keyword spotting experiment using perceptually significant features

This paper presents the preliminary results of work carried out for recognizing certain keywords using perceptually significant spectral energy features. Dynamic time warping and artificial neural networks were used for feature matching. Preliminary results indicate that the significant energy features are feasible as a stand-alone set that can also augment the most commonly used cepstral features to yield high recognition scores. For the challenging set of short words used in the present work, results show that a neural network for feature recognition is better than a dynamic time warping technique with different dissimilarity measures.