Speaker Recognition UsingPulse Coupled Neural Networks
Antonio PedroTimoszczuk · 2007
PulseCoupledNeuralNetwork(PCNN)isa paradigm thathasnotyetbeenexplored enoughinspeaker recognition. Thispaperpresents theresults ofexperiments conducted todevelop a new recognition architecture that applies PCNN totextindependent speaker recognition. The proposed architecture comprises atwolayer PCNNforfeature extraction and a Multilayer Perceptron (MLP)forfinal classification. Thefirst layer ofPCNNperforms apulse coding taskanditsoutputs areusedasinputs toa secondself- organizing layer thatlearns thepulse's statistics. Initially, the wellknownMel-Cepstral Frequency Coefficients wereusedasa benchmarktoverify thenew architecture's capability of learning thespeaker information. After these preliminary tests, theMel scaled filter bankenergycoefficients wereused, resulting ina biologically plausible architecture. Thefirst results demonstrated thePCNN'sability todealwithtemporal information contained in speechsignals. The proposed architecture ispromising, presenting 82% recognition when compared with96% oftheclassical MLP classifier.