Fred Cummins
Depts. of Linguistics and Cognitive Sci., Indiana Univ., Bloomington, IN 47405
The problem of pattern recognition in time is usually addressed by buffering, which converts time into a spatial dimension, and allows the application of standard pattern recognition methods. But buffering of high bandwidth sensory input is implausible and problematical. An architecture based on the adaptive oscillator model [J. D. McAuley, J. Acoust. Soc. Am. 95, 2966 (A) (1994)] is presented which generates a spatial pattern from the rhythmic content of the acoustic input. The acoustic signal is passed through a bank of gammatone filters, each channel is half-wave rectified and down sampled, allowing a simple differencing procedure to identify onsets, which serve as inputs to a 2-D array of oscillators organized by frequency channel and by intrinsic period. Each oscillator adjusts its intrinsic period to match periodic onsets present in the signal. Only those oscillators that succeed in synchronizing their activity to a period in the input signal produce persistant output. After synchronization, the output distribution produces a stable spatial pattern over the 2-D array. The procedure allows treatment of pattern distributed in time without recourse to sensory buffering. Output patterns for a variety of stimuli, including animal gaits, musical rhythms, and prosodic structure will be presented. [This project was supported by ONR.]