Using spiking onset neurons and a recurrent neural network for musical sound classification

Citation

Newton M & Smith L (2011) Using spiking onset neurons and a recurrent neural network for musical sound classification. 161st Meeting of the Acoustical Society of America, Seattle, WA, USA, 23/05/2011 - 27/05/2011. Journal of the Acoustical Society of America, 129 (4 part 2), pp. 2486-2486. http://scitation.aip.org/getpdf/servlet/GetPDFServlet?filetype=pdf&id=JASMAN000129000004002486000003&idtype=cvips&prog=normal

Abstract
Physiological evidence suggests that specific neurons within the cochlear nucleus specialize in sound onset detection. These are innervated by type 1 spiral ganglion fibers covering a relatively wide spectrum. Sudden increases in sound energy e.g., during the initial portion of a sound result in an increased firing rate in a downstream onset neuron. Onset timing and spectral location are thought to play a role both in auditory stream separation and sound identification and interpretation. Onset neurons are modeled using leaky integrate-and-fire units innervated by spiking data streams produced using a passive gammatone filterbank followed by positive-going zero-crossing detection. Signal level is coded using multiple spike trains per filter channel. The model is presented with a succession of 607 musical samples selected from the McGill dataset and the pattern of onset spikes recorded for each sound. Groups of onset spikes occur close to the beginning of each note. The objective is to use the pattern of spikes, produced by the onset neuron model, as a fingerprint of the original acoustic signal. These onset fingerprints are presented to a recurrent neural network reservoir network to attempt to classify them. The results are compared with a sound classification scheme based on cepstral coefficients.

Journal
Journal of the Acoustical Society of America: Volume 129, Issue 4 part 2