ABSTRACT
This chapter explains with a feature-extraction approach to auditory pattern recognition. It presents a comparative study of pattern recognition by humans and pattern recognition by machine. Cluster-seeking and feature-extraction approaches are proposed for automatic recognition of auditory patterns. The pattern-recognition process may be modeled as a mapping process that consists of three spaces: measurement space, feature space, and category space. Auditory, nonspeech signals are generally characterized by loudness, pitch, duration, and timbre as primary perceptual attributes. In automatic classification by cluster seeking, people usually encounter two basic problems. The first is the determination of an optimal number of cluster centers; the second is the assignment of patterns into identified clusters. In the case of machine recognition, discriminating attributes are extracted to form a feature vector or a feature string. Examples of the kinds of features are statistical means, correlation coefficients, eigenvalues, eigenvectors of covariance matrices, Fourier coefficients, and other invariant properties.
