A BI-DIRECTIONAL MODEL OF ENGLISH PRONUNCIATION

The two tasks of finding the pronunciation of a word from its spelling, and the spelling from its pronunciation, are basic problems in speech synthesis and recognition respectively. A related problem is how best to align the phonemic and orthographic representations of a given word to show the correspondence between each of the letters in the word and the sounds to which they belong. The finding of one form of a word when observing only the other form, is likened to decoding an encrypted message to find a hidden meaning. If a Hidden Markov Model (HMM) is assumed to generate the observed form of the word from its hidden form, then a method exists to solve the alignment problem, provided that the parameters of the model are known. Since they are in general not accurately known, a training algorithm, such as the Forward-Backward (maximum-likelihood) method can be used to determine a good estimate for them. A simple HMM for solving the two decoding tasks is suggested, and the results of training it on real data are discussed. The use of a single methodology to solve two different but related tasks is offered as an example for other language tasks.