Isolated spoken number recognition with hybrid of self-organizing map and multilayer perceptron
Perttu Salmela, Kari Laurila, Seppo Kuusisto, Petri Haavisto, Jukka P. P. Saarinen · 2002
A neural network, which is capable of recognizing isolated spoken numbers speaker independently is described. The recognition system is a hybrid of a self-organizing map (SOM) and a multilayer perceptron (MLP). The SOM maps the feature vectors of a word in a constant dimension binary matrix, which is classified by a MLP. The decision borders of the SOM were fine-tuned with the LVQ1 algorithm, with which the hybrid achieved over 99% recognition out of 1232 test set samples. The training convergence of the MLP was tested with two different initialization methods.