Dynamic neural network for speech recognition

Dynamic neural network for vocal signal recognition is proposed. The neural network forms the speech image templates in the manner of networks of states and transitions while covering the signal mapping area in signal space by a multitude of interconnected hyperspheres. Network training is realized by dynamic separation of neurons using an algorithm Kohonen’s self-organizing mappings. Recognition is considered as a dynamic process controlled by attention mechanism. The results of some experiments with neural network model are described.