A hybrid neural network/rule based architecture used as a text to phoneme transcriber

A major stage in the synthesis of natural sounding speech from unrestricted text is the transcription from normalised text into sound representative code. Previously research has either concentrated on the improvement of rule based algorithms, or has used a very large scale neural network to carry out this task. A hybrid neural network and simplified rule base system has been proved to simulate the MOSFET operation more efficiently than a neural network alone. This paper describes the extension of this idea to the text to phoneme transcription problem. Optimum neural network size and configuration and rule base simplifications are investigated