A template-based approach for recognition of intermittent sounds

Abstract

Automatic speech and sound recognition typically involves some measure of distance between training and (possibly time-warped) test samples. Special problems arise when the spectral samples of interest are intermittent and contain temporal patterns of alternating periods of sounds and pauses that are significant for recognition. In such cases a recognizer must be capable of distinguishing between the end-points and the pauses of digitized samples and economically searching the segmented sounds for the occurrence of significant spectral patterns. The usual distance metrics based on conventional dynamic time warping algorithms may be inappropriate because time-warping often corrupts the temporal structure of the sound. The problem can be solved by first searching a test sample for distinctive temporal patterns and, if more than one match is obtained, using a spectral distance measure to classify the sample with its nearest neighbor among these. Computational advantages can be obtained if both the temporal and spectral templates are maintained in a binary format reflecting the important sound components.