Statistical Grapheme to Phoneme Conversion using Language Origin

This report describes a method for grapheme to phoneme conversion using statistical models of pronunciation. The available techniques for this conversion are first described and examples of each are given. A baseline system which uses Hidden Markov Models to represent phonemes in English is described and evaluated. The results from the baseline system serve to replicate previous research and to function as a means of comparison for extensions created for this project. The system is first extended to multiple languages using the CELEX English, German and Dutch lexicons, and results of 40.89%, 78.55% and 74.48% words correct respectively are obtained for a 4-gram phoneme model. The report then describes the classification of language origin using a letter n-gram model, similar to the language model in an automatic speech recognition system. An implementation of the language recogniser gives an overall language recognition accuracy of 90.48%. The language identification and grapheme to phoneme stages are combined to create an integrated system. Two general methods of combining the stages are investigated. The first involves the language identification stage generating a result about the language origin of a word and the grapheme to phoneme converter appropriate to that language being used. The second approach combines the language identification probabilities with the values returned by the grapheme to phoneme models for all the languages. It is found that such a probabilistic treatment of language origin provides an enhanced result compared with a hard decision by the language classifier.