Improving Pronunciation Accuracy of Proper Names with Language Origin Classes

Pronunciation of proper names that have different and varied language sources is an extremely hard task, even for humans. This thesis presents an attempt to improve automatic pronunciation of proper names by modeling the way humans do it, and tries to eliminate synthesis errors that humans would never make. It does so by taking into account the different language and language family sources and by adding such information as features into the pronunciation models, either directly or indirectly. This approach does result in an improvement of pronunciation accuracy, however in order to assess the true goodness of this approach, we would need to develop a more accurate language identifier. Ultimately, the data we would like to have in order to train our models is a list of proper names tagged both with their phonetic transcription and with the language they come from. A new approach this thesis begins to investigate is the unsupervised clustering of proper names to derive language classes in a data-driven way. With this approach, no language classes (Catalan, English, French, German, etc.) need to be determined a priori, but rather they are inferred from the names and their pronunciation. The clustering method used takes into account letter trigrams as well as their aligned pronunciation at training time. Experiments using the classes derived from unsupervised clustering are still preliminary and have not yet yielded an improvement in pronunciation accuracy of proper names.