A Diagonally Weighted Binary Memristor Crossbar Architecture Based on Multilayer Neural Network for Better Accuracy Rate in Speech Recognition Application

Abstract

A novel binary memristor crossbar architecture based on multilayer neural networks is proposed in the speech recognition application. Here, the memristor crossbar circuit acts as the weights of the neural network combined with the activation function circuit to determine the output. In the new crossbar architecture, the weights are arranged diagonally and divided into 2 arrays according to positive and negative weights. A speech recognition application for 5 vowels is implemented using the proposed architecture. The result shows that the average recognition rate achieves from 94 percent to 96.6 percent over 1000 audio samples. A statistical table shows that the recognition rate and the number of the memristors increase correspondingly to the number of used bits. From the Monte Carlo simulation, the recognition rate of the proposed binary memristor crossbar is decreased slightly from 94 percent to 93.7 percent, while the memristance variation is increased from 1 percent to 15 percent.

Keywords

• pattern recognition
• memristors
• neural network
• neural network hardware
• speech recognition