Adapting automatic speech recognition for foreign language learners in a serious game

Abstract

EveiBd is a project for development of an immersive, virtual "serious game" for computer assisted foreign language learning, with which users interact verbally by means of an automatic speech recogniser. The speech of the target user group, namely adolescent low-proficiency non-native speakers, differs significantly from native adult speakers on which ASR systems are typically trained and thus on which they perform the best. As only a very small corpus of near-domain speech and text is available, the aforementioned difference becomes a development challenge. We deal with this challenge by adapting existing systems on the small data set. We adapt our language models using text selection to augment the in-domain data with similar data from out-of-domain sources. We adapt our acoustic models with MAP adaptation. Through these steps we achieve significant reductions in error.