Previous work has demonstrated that multilayer perceptrons may be used to transcribe orthographic text into a phonemic representation. It is shown that this technique may be applied to a large dictionary of over 70,000 words without reduction in performance. The dictionary was divided into sub-dictionaries containing regularly and irregularly pronounced words using a rule-based system. The multilayer perceptron performed better with the regularly pronounced words. However, combining two separately trained networks resulted in little better performance than a single network trained on the whole dictionary.
Bibliographic reference. Ainsworth, William A. / Pell, B. (1989): "Connectionist architectures for a text-to-speech system", In EUROSPEECH-1989, 1125-1128.