SRI'S 2004 BROADCAST NEWS SPEECH TO TEXT SYSTEM
Anand Venkataraman, Raghudeep Gadde, Andreas Stolcke, Dimitra Vergyri, W. Wang, Jing Wen Zheng · 2004
We describe the system that we used to automatically transcribeBroadcastNewsSpeechin the2004NIST evaluationssponsored by the DARPA EARS program. We also describe some of our efforts leading up to the NIST evaluation, in particular, concerning data selection and model training. Since one of the hallmarksofthe 2004evaluationwasthe availabilityof relatively large amounts of quickly transcribed data, we describe our techniques, and those of others in adapting this data for model training, and experiments attempting to evaluate the suitability of this data for training purposes. We concludeby presenting results and also discussing whether or not increases in amounts of training data proportionately benefit the research community.