Towards lecture transcription in resource-scarce environments

Abstract

We present progress towards automated Lecture Transcription (LT) in resource scarce environments. Our development has focused on the transcription of lectures in Afrikaans from two faculties at North-West University. A bootstrapping procedure is followed to filter and select well-aligned segments of speech. These segments are then used to train acoustic models. Initial work towards language modeling for LT in a resource-scarce environment is also presented; manual lecture transcriptions are combined with text mined from other sources such as study guides to train language models. Interpolation results indicate that study guides are a useful resource for language modeling, whereas general text (obtained from a publisher of Afrikaans books) is less useful in this context. Our findings are confirmed by the reduced word error rates (WERs) obtained from our off-line speech-recognition system for Lecture Transcription.