Hardware radial basis functions neural networks for phoneme recognition

Abstract

The ability of a neural network to learn on-line is crucial for real time speech recognition systems. In fact, analog neural network systems are preferred to their digital counterparts mainly due to the high speed that they can attain. However, the training method adopted also affects the performance of the neural network. The conventional error backpropagation network usually requires quite a long convergence time for correct weight adjustment since the sigmoid function of a conventional multilayer network gives a smooth response over a wide range of input values. In contrast, the Gaussian function responds significantly only to local regions of the space of input values. Thus, backpropagation training is more efficient in neural networks based on Gaussian functions or radial basis function (RBF) networks, than those based on sigmoid functions in the hidden layer. The paper proposes an analog VLSI chip, which can be cascaded in order to develop an RBF neural network system for phoneme recognition.