14th International Congress of Phonetic Sciences (ICPhS-14) San Francisco, CA, USA August 1-7, 1999

A Multilingual Acoustic-Phonetic Model based on Phone-to-Phone Transition Classes

Giorgio Micca (1), Enrico Palme (2), Juan Mari (3)

(1) CSELT, Turin, Italy; (2) Università di Pisa, Italy;
(3) Universidad Politècnica de Catalunya, Barcelona, Spain

The paper presents an acoustic-phonetic-model based on classes of transitions between phones. The model is to be used in the development of speech recognition systems for multilingual applications. The problem that has been tackled is twofold: how to deal with application vocabularies containing multiple language entries and how to cope with scarce training material in a new target language. The Italian, Spanish and British English languages have been considered in this work. Monolingual models, based on phone transition classes, show recognition performance comparable or even better to alternative models requiring larger inventories of units. Results for a three-lingual model, consisting of a global set of acoustic-phonetic classes where similar transition sounds for the three languages have been unified, are also reported. The approach is being experimented in “blind” mode with the Rumanian language, where the phonological structure of this language is the only information source feeding the recognition model.

Full Paper

Bibliographic reference.  Micca, Giorgio / Palme, Enrico / Mari, Juan (1999): "A multilingual acoustic-phonetic model based on phone-to-phone transition classes", In ICPhS-14, 1677-1680.