RECENT PROGRESS IN THE DECODING OF NON-NATIVE SPEECH WITH MULTILINGUAL ACOUSTIC MODELS

摘要

In this paper we report on recent progress in the use of multilingual Hidden Markov Models for the recognition of non-native speech. While we have previously discussed the use of bilingual acoustic models and recognizer combination methods, we now seek to avoid the increased computational load imposed by methods such as ROVER by focusing on acoustic models that share training data from 5 languages. Our investigations concentrate on the determination of a proper model complexity and show the multilingual models' capability to handle cases where a non-native speaker is borrowing phones from his or her native language. Finally, using a limited amount of non-native speech for MLLR adaptation, we demonstrate the superiority of multilingual models even after adaptation.