Linguistic Resources for Meeting Speech Recognition

This paper describes efforts by the University of Pennsylvania's Linguistic Data Consortium to create and distribute shared linguistic resources – including data, annotations, tools and infrastructure – to support the Rich Transcription 2005 Spring Meeting Recognition Evaluation. In addition to distributing large volumes of training data, LDC produced reference transcripts for the RT-05S conference room evaluation corpus, which represents a variety of subjects, scenarios and recording conditions. Careful verbatim reference transcripts including rich markup were created for all two hours of data. One hour was also selected for a contrastive study using a quick transcription methodology. We review the two methodologies and discuss qualitative differences in the resulting transcripts. Finally, we describe infrastructure development including transcription tools to support our efforts.