A fast text-to-speech learning technique is presented that takes into account the maximum information that is available to the reader for each word. This formulation is based on the connectionist Optimum Path Paradigm and is consistent with recent theories of "natural" phonology. It is capable of capturing allophonic variations in a variety of languages with the minimum needed retraining. The procedure is described and completed by the specification of graph traversal and learning of the W phoneme- transition weight matrix which is related to a Hebbian form of learning. Results for learning text-to-speech for several languages as well as implementation aspects are discussed.