Adapted speech recognizer to improve English pronunciation of Hispanics

Abstract

This work presents advances on an adapted speech recognition system, based on hidden-Markov models (HMMs), to help Hispanics in their pronunciation of English. The main results so far, with English digits, show significant recognition improvements of the adapted recognizer over the not adapted one, both for one speaker (100% vs 94%) and for groups of them (96% vs 74%). The adaptation uses maximum likelihood linear regression (MLLR). Two databases were used in the experiments. The TIMIT database to train the recognizer models with English native speakers, and the Latin-American Spanish database to adapt and test the adapted and not adapted recognizers. The project, of which this work is a part, is an English pronunciation tutor, whose main parts are this adapted speech recognizer, a pronunciation evaluator (phonetic and prosodic) and a dialog manager. English speech recognition systems are trained with native speakers, and most of the recognition errors are attributed to the interference of foreign accents. This is why it is necessary to adapt the models to take into account the characteristic features of a given population; in this case Hispanics. (To be presented in Spanish.)