Maximizing the target-pattern cross-correlation for training time-delay neural networks

In this paper experimental conclusions are reported on the verification of a new learning procedure for training time delay neural networks (TDNN), based on the maximization of the cross-correlation between the output of the network (pattern) and the target reference sequence. This functional has been used for training a TDNN encharged of estimating the aperture of the speaker´s mouth from the acoustic analysis of his speech. Performances have been compared to those reported in a previous paper obtained with classical MSE-based back-propagation. Experimental results provide clear evidence of the improvements, both in terms of convergence speed and estimation fidelity, achievable through this new training algorithm