A Cortically-Inspired Model for Bioacoustics Recognition

Wavelet transforms have shown superior performance in auditory recognition tasks compared to the more commonly used Mel-Frequency Cepstral Coefficients, and offer the ability to more closely model the frequency response behaviour of the cochlear basilar membrane. In this paper we evaluate a gammatone wavelet as a preprocessor for the Hierarchical Temporal Memory (HTM) model of the neocortex as part of the broader development of a biologically motivated approach to sound recognition. Specifically, we apply for the first time, a gammatone/equivalent rectangular bandwidth wavelet transform in conjunction with the HTM's Spatial Pooler to recognise frog calls, bird songs and insect sounds. Our audio feature detection results show that wavelets perform considerably better than MFCCs on our selected datasets but that combining wavelets with HTM does not produce further improvements. This outcome raises questions concerning the degree of match to the biology required for an effective HTM-based model of audition.