A method is presented that detects behavioral transients in waveforms. A structure is defined that accepts waveforms as inputs and generates a sequence of symbols representing the sequence of transients present in the waveform. This structure is developed by generalizing an unsupervised learning algorithm to the time-varying case. The algorithm accepts a sequence of unlabeled waveforms to find cluster centers associated with the transients. Clustering is assumed to be with respect to an arbitrary distance measure. This measure is assumed to satisfy differentiability and regularity requirements. The algorithm is shown to converge based on assumptions concerning a unique optimum. This is done by the application of a stochastic-approximation theorem to a gradient-following technique. The resulting algorithm is applied to a problem in speech processing. The structure resulting from the learning algorithm is compared to the standard linguistic phonetic structure.