Language Model Adaptation with the Use of Presentation Slide Information for Automatic Lecture Transcription

We propose a language model adaptation method with the use of presentation slide information for automatic lecture transcription. N-gram probabilities are rescaled with lecture-dependent unigram probabilities estimated by PLSA using all slides of the lecture. In addition, the N-gram language model is interpolated with a model trained with the Web texts collected via the Web search, using keywords extracted from the slides. Moreover, N-best hypotheses of ASR are rescored using word probabilities enhanced with a cache model using the slide corresponding to each utterance. Experimental evaluations on real lectures show that the proposed method with the combination of the global and local slide information achieves a significant improvement of ASR accuracy.