A trainable voice recognition ic based on the human auditory cortex

In this paper, at first a new topology of neural network is introduced. This network consists of four modular layers which recognize the frequencies of the human voice. The first layer of the network eliminates frequencies with weak intensities. The second layer recognizes the pitch or the first harmony of the voice. The third layer prepares inputs based on other harmonies of the frequency for the last layer which is an associative memory neural network to map the input set to a desire output set. At the end, an integrated circuit which is based on trainable transistors derived from the associative memory layer as the most important and complicated part of the network is presented. Each trainable transistor has two wires for trapping or removing electrons in or from a floating gate. This new structure makes a fast trainable IC.