Feature Histograms as a Model of Speech Perception

This paper reports the results of an investigation into the use of a discrete set of MFCC features in a speaker-independent. isolated-word recogniser.and discusses the implications for speech perception.In these experiments.a discrete set of feature types is obtained by Vector Quantising frames of speech encoded as Mel-frequency Cepstral Coef cients (MFCC).Since the frames of speech from which the MFCCs are derived are over-lapped to a large extent, sequence information is contained.implicitly. in the frames which occur.Thus an utterance may be represented by a simple histogram of the feature-types it contains.The frame duration and overlap in this scheme, are optimised using a simple recognition paradigm.In training mode, a feature histogram.which acts as a template, is constructed for each word in the vocabulary.In recognition mode.the histogram of the unknown weird is compared with the template histograms.