Neural timing nets for fundamental frequency (F0)-based auditory scene analysis.

Abstract

Neural timing nets are idealized networks of delay lines, coincidence detectors, and adaptive processing elements that operate in the time domain on temporally coded signals to compare, extract, and separate auditory objects [Cariani, Neural Networks 14, 737–753 (2001); J. New Music Res. 30, 107–135; IEEE Trans. Neural Netw. 15(5) (2004)]. Timing nets constitute an alternative, potential mode of neural signal processing in which information resides in neural signals rather than in patterns of activated elements. Recurrent timing nets with delay loops act as dense arrays of recursive, comblike filters to effect a period-by-period analysis that builds up and separates component invariant time patterns with different fundamentals (F0s). Using both linear and nonlinear processing rules, the latter were used to process and separate synthetic double vowels, running speech, and polyphonic musical excerpts, with varying results. Relations to, and combinations with, processing strategies based on autocorrelation and all-order interspike intervals, adaptive comb filtering, correlogram duplex analysis, cancellation, Fourier zero-crossings, and demodulation, with and without prior bandpass filtering, are discussed. Bottom-up/top-down mechanisms for dynamic facilitation of lower level temporal processing loops are also considered. [Work supported by AFOSR FA9550-09-1-0119.]