Sequential feature extraction for waveform recognition

A means of feature extraction and recognition for waveforms is developed and applied to speech recognition. The concept of sequential feature extraction is formalized and a performance criterion for the resulting extraction is developed. An unsupervised learning algorithm, which will optimize this structure with respect to the performance criterion, is presented. This algorithm, which can be applied to waveform recognition as well as vector recognition, represents an improvement over existing clustering algorithms in many respects. This method will allow unbounded stringsof sample patterns for learning. The samples are presented to the algorithm one at a time so that the storage of large numbers of patterns is unnecessary. The assumption of known probability measures is not made.