A hybrid neural network/rule based architecture for diphone speech synthesis

Analogue neural networks (ANNs) have successfully been applied to controlling a formant speech synthesiser, resulting in high quality speech. However they are somewhat limited by the large number of hidden layer neurons needed. The paper describes the application of a hybrid ANN/rule-based optimised computing architecture to diphone speech synthesis. The architecture utilises a simplified rule-base, based on a diphone data base, and an ANN working in parallel. The number of hidden layer neurons in the ANN unit when used in parallel with the rule-base is reduced when compared to the hidden layer size of a standalone ANN used for diphone synthesis. This reduction in hidden layer size results in faster learning, with no reduction in overall system performance being observed