TR-SLT-0093 :2005.03.31

Rainer Gruhn, Konstantin Markov

Pronunciation Modeling with HMMs as Statistical Lexica

Abstract:Non-native speakers pronounce words in multiple different ways than native speakers. To model these deviations statistically, we propose discrete word HMMs as statistical lexicon. The models are initialized from some baseline lexicon and trained on the result of phoneme recognition. The models are applied by rescoring an n-best recognition result.