Human Language Technologies (HLT) – Pronunciation modelling

Given the spelling of a word in a specific language, what would that word sound like? A pronunciation model describes this process of letter-to-sound conversion. A pronunciation modelling component is required by many speech processing tasks – including general domain speech synthesis and large vocabulary speech recognition – and is often one of the first resources required when developing speech technology in a new language.

The letter-to-sound relationship can either be modelled through explicit pronunciation dictionaries or abstract letter-to-sound formalisms; the latter can be used to predict the pronunciation of words that have not been seen before. These more general grapheme-to-phoneme relationships can be derived directly from a training dictionary using techniques such as decision trees, pronunciation-by-analogy (PbA), or our personal favourite: the Default&Refine algorithm, a very efficient algorithm when training on small dictionaries.

Our research focuses on the fast and efficient development of pronunciation dictionaries in resource-scarce languages, and includes:

• The development of DictionaryMaker for the rapid bootstrapping of pronunciation dictionaries.
• Efficient techniques for error detection in pronunciation dictionaries.
• Adapting dictionaries to different dialects.
• The development of multilingual dictionaries in code-switching environments (where different languages are mixed in a single sentence).

The multilingual environment of South Africa provides the ideal smorgasbord of languages and dialects within which to conduct pronunciation research: not just about each language individually but also about the influence that the various languages have on one another.

Links

DictionaryMaker: Tool for the rapid bootstrapping of pronunciation dictionaries.

Selected Publications

M. Davel and E. Barnard, “Pronunciation prediction with Default&Refine”, Computer Speech and Language, Vol 22, pp 374-393, October 2008.

O.M. Martirosian and M. Davel, “Error analysis of a public domain pronunciation dictionary”, In Proceedings of the 18th Annual Symposium of the Pattern Recognition Association of South Africa (PRASA), pp 13-16, Pietermaritzburg, South Africa, November 2007.

M. Davel and E. Barnard, “Effort and Accuracy during Language Resource Generation: A Pronunciation Prediction Case Study”, SAIEE Africa Research Journal, Vol 98 (4), pp 124-128, December 2007.

M. Davel and E. Barnard, “Developing consistent pronunciation variants”, Interspeech, Pittsburgh, USA, September 2006.

M. Davel and E. Barnard, “The Efficient Generation of Pronunciation Dictionaries: Human Factors during Bootstrapping”. In Proceedings of the 8th International Conference on Spoken Language Processing, Korea, 2004.