Hiertalker: A default hierarchy of high order neural networks that learns to read english aloud

It is a problem of general interest to determine the relationships among two or more sets of discrete symbols. This problem occurs in translating languages, writing compilers for computer languages, determining the correspondence between genetic codes and their structural realizations, and in mapping ordinary spelling onto a phonetic transcription appropriate to drive a speech synthesizer. We proposed and tested a new learning procedure, based on a default hierarchy of high order neural networks, which exhibited an enhanced capability of generalization and a good efficiency. This new architecture is suitable for learning regularities embedded in a stream of information with inherent long range correlations. When applied to the conversion of English words to phonemes, a simulator of such a hierarchy, HIERtalker, achieved an accuracy of typically 99% for the words in the training set, and 96% for new words. Also, HIERtalker used considerably less computer time than NETtalk did.