Application of dynamic synapse neural networks on identification and localization of nonspeech sounds
Alireza Afshordi Dibazar, Theodore W. Berger · The Journal of the Acoustical Society of America · 2004
This paper focuses on the dynamic synapses neural network (DSNN) for identification of nonspeech sounds, including the chambering of a gun, as well as localization of identified sounds. The algorithm employed consisted of extracting DSNN features from sounds and classification of features based on Gaussian mixture models (GMMs). To extract DSNN features, a single neuron with a presynapse including a 14th order differential equation, first order post-synapse, and first order inhibitory feedback was used. After training, network parameters were used as features. The classification task was then formulated as an estimation of conditional joint probability. Classification results were collected from both chambering identification and localization. Successful localization was defined as correct identification of speaker of origin. The number of training and testing samples were 120 and 160, respectively. The system performed 96.88% and 90.00% correct identification and localization of test samples.