Neural networks with random letter codes for text-to-phoneme mapping and small training dictionary

In this paper we address the problem of text-to-phoneme (TTP) mapping implemented by neural networks. One important disadvantage of the neural networks is the convergence interval which can be in some situations very large. Even when the neural networks are trained in off line mode a shorter convergence interval would be of interest due to various reasons. In the TTP mapping, decreasing the number of necessary iterations is equivalent to relaxing the requirements for the dictionary size. In this paper, we show that proper letter encoding can increase the convergence speed of the multilayer perceptron neural network for the task of TTP mapping. Experimental results that compare the performance of several techniques that speed-up the convergence of the multilayer perceptron, in the context of TTP mapping are also presented.