Improving Automatic Speech Recognition for Lectures through Transformation-based Rules Learned from Minimal Data

Proceedings of the 47th Annual Meeting of the ACL and the 4th IJCNLP of the AFNLP, pages 764–772, Suntec, Singapore, 2-7 August 2009. c⃝2009 ACL and AFNLP Improving Automatic Speech Recognition for Lectures through Transformation-based Rules Learned from Minimal Data Cosmin Munteanu∗† ∗National Research Council Canada 46 Dineen Drive Fredericton E3B 9W4, CANADA Cosmin.Munteanu@nrc.gc.ca Gerald Penn† †University of Toronto Dept. of Computer Science Toronto M5S 3G4, CANADA {gpenn,xzhu}@cs.toronto.edu Xiaodan Zhu† Abstract We demonstrate that transformation-based learning can be used to correct noisy speech recognition transcripts in the lec- ture domain with an average word error rate reduction of 12.9%. Our method is distinguished from earlier related work by its robustness to small amounts of training data, and its resulting efficiency, in spite of its use of true word error rate computations as a rule scoring function. 1 Introduction Improving access to archives of recorded lectures is a task that, by its very nature, requires research efforts common to both Automatic Speech Recog- nition (ASR) and Human-Computer Interaction (HCI). One of the main challenges to integrating text transcripts into archives of webcast lectures is the poor performance of ASR systems on lecture transcription. This is in part caused by the mis- match between the language used in a lecture and the predictive language models employed by most ASR systems. Most ASR systems achieve Word Error Rates (WERs) of about 40-45% in realis- tic and uncontrolled lecture conditions (Leeuwis et al., 2003; Hsu and Glass, 2006). Progress in ASR for this genre requires both better acoustic modelling (Park et al., 2005; F¨ugen et al., 2006) and better language modelling (Leeuwis et al., 2003; Kato et al., 2000; Munteanu et al., 2007). In contrast to some unsupervised ap- proaches to language modelling that require large amounts of manual transcription, either from the same instructor or on the same topic (Nanjo and Kawahara, 2003; Niesler and Willett, 2002), the solution proposed by Glass et al. (2007) uses half of the lectures in a semester course to train an ASR system for the other half or for when the course is next offered, and still results in signifi- cant WER reductions. And yet even in this sce- nario, the business case for manually transcrib- ing half of the lecture material in every recorded course is difficult to make, to say the least. Manu- ally transcribing a one-hour recorded lecture re- quires at least 5 hours in the hands of qualified transcribers (Hazen, 2006) and roughly 10 hours by students enrolled in the course (Munteanu et al., 2008). As argued by Hazen (2006), any ASR improvements that rely on manual transcripts need to offer a balance between the cost of producing those transcripts and the amount of improvement (i.e. WER reductions). There is some work that specializes in adap- tive language modelling with extremely limited amounts of manual transcripts. Klakow (2000) filters the corpus on which language models are trained in order to retain the parts that are more similar to the correct transcripts on a …

Content truncated to 500 words