The Accurate Estimation of Articulatory Synthesiser Parameters through Reducing the Degree of Saturation

A new method is proposed to correctly estimate the parameters of an articulatory speech synthesiser using a MLP neural network. This is achieved through modifying the statistical characteristic of the acoustic input pattern vectors in order to prevent the activation level of the hidden nodes from approaching saturation. The technique results in considerably faster neural learning and a more accurate estimation of the articulatory synthesiser parameters.