Knowledge of language origin improves pronunciation accuracy of proper names

As it is impossible to have a lexicon with complete coverage, and a high proportion of unknown words are proper names, this paper addresses the issue of automatically finding pronunciations of unseen proper names in US English. Proper names, especially in the US, may come from a large range of ethnic backgrounds. We present a model and results showing that including ethnic origin of words in a statistical model can improve pronunciation results. We used a lexicon of 56,000 proper names from CMUDICT. We also gathered data (text and proper names) from 26 languages to built statistical models that provide an estimate of word origin. Tests against held out data showed a 7.6% absolute improvement from a baseline of 54.8% when language based features were added to our CART-based model. As there are potentially multiple correct pronunciations, we synthesized a random sample of names that did not match the “correct” answer in our test set. Human listeners showed a 17% preference for the model with language features compared to the baseline.