Learning of word stress in a sub-optimal second order back-propagation neural network

The authors show an example of an efficient and easy solution, using a neural network, of a problem that cannot be easily solved with rules. This example regards the localization of primary word stress. The knowledge of the position of primary stress is very useful in text-to-speech synthesis of Italian, a language characterized by a very prominent word accent. In fact, the position of word stress is the basis for the automatic generation of the pattern of duration of the syllables and of the intonation of the whole phrase. The authors use a feedforward network with an error backpropagation learning, extending the method with the computation of the correction step based on the second derivative of the error function. This method has been used to speed up convergence without using a fixed learning rate and a momentum term. The authors obtain a steep decrease of the error at the expense of a limited increase of the computational cost.<>