Description:
Neuroscience shows that auditory neurons extract a variety of specific signal parameters such as frequency, amplitude modulation, frequency modulation, and onsets. Furthermore, biology gives us examples of auditory foveae in bats and owls where there is overrepresentation of behaviorally relevant signal parameters. This shows that biology finds nuances of these parameters to be more important than compact coding considerations such as orthonormal basis functions. In contrast to typical spectrogram approaches where a specific time window is chosen and where the Fast Fourier Transform provides non-overlapping orthogonal rectangular tilings, we start with the smallest tiling possible. The uncertainty principle gives us this tiling as the Gaussian modulated sinusoid. We can rotate this tiling in the time-frequency plane, which gives us a chirplet. We then construct many chirplets (a chirplet set) centered at the same time-frequency point. We then find the chirplet that best matches the signal at this point. Our ability to sample at arbitrary locations in the time-frequency plane with different chirplet sets gives us the ability to create customizable artificial auditory foveae. We use this chirplet front-end for a classification task of identifying Marmoset vocalizations and comparisons to typical spectrogram methods for audio classification are made.