ISCA Archive Eurospeech 1999
ISCA Archive Eurospeech 1999

Further advances in transcription of broadcast news

Long Nguyen, Spyros Matsoukas, Jason Davenport, Daben Liu, Jay Billa, Francis Kubala, John Makhoul

In this paper, we describe our recent significant progress in automatic transcription of broadcast news programming from radio and television, compared to that described in [1] at EuroSpeech97. Overall, we achieve a 42% relative word error rate reduction during the latest DARPA November 1998 Hub-4 Evaluation, in contrast to the 1996 evaluation result reported in [1]. This significant progress was attributed to effort in improvements of segmentation, spontaneous speech modeling, careful processing of acoustic training data, bandwidth-specific acoustic models, as well as other improvements in the basic acoustic modeling field.