Approaches to foreign-accented speaker-independent speech recognition

Current research in the area of foreign-accented speech recognition focusses either on acoustic model adaptation or speakerdependent pronunciation variation modeling. In this paper both approaches are applied in parallel and in a speaker-independent fashion: the acoustic modeling part is based on a derived Hidden Markov Model (HMM) clustering algorithm and the lexicon adaptation is based on speakerindependent multiple pronunciation rules. The pronunciation rules are derived using phoneme-level pronunciation scores. Foreign-accented speech was simulated with Columbian Spanish and Spanish of Spain and the experiments showed an improved recognition performance for the acoustic modeling part and identical recognition results when adding pronunciation variants to the lexica. Both results are taken as indicators for an improved recognition performance when applied on real foreignaccented speech. The present limited availability of foreignaccented speech databases, however, clearly merits further investigations.