The 1998 BBN Byblos 10 x Real Time System

In this paper we describe the BBN Byblos 10x real t ime system used for the 1998 Hub-4 English tests. Given our s tate of the art primary system [1] running at 230 times real time ( 230 xRT) we show that eliminating and approximating many comput ationally expensive components speeds up the system by a fact or of 23 with a relative loss in WER of 18%. This is accomp lished without retraining or changing the primary system s tructure. The components of the primary system that are refined i nclude segmentation, adaptation, decoding, cross-word resc o ing with adaptation, and system combination. The time savin g algorithms used include fast Gaussian computation, grammar spr eading, nbest tree rescoring, and block diagonal adaptation .

[1]  John Makhoul,et al.  The 1998 BBN BYBLOS Primary System applied to English and Spanish Broadcast News Transcription , 1998 .

[2]  Richard M. Schwartz,et al.  Efficient 2-pass n-best decoder , 1997, EUROSPEECH.

[3]  J. Makhoul,et al.  Vector quantization in speech coding , 1985, Proceedings of the IEEE.

[4]  Bruce T. Lowerre,et al.  The HARPY speech recognition system , 1976 .

[5]  Michael Picheny,et al.  Decision-tree based feature-space quantization for fast Gaussian computation , 1997, 1997 IEEE Workshop on Automatic Speech Recognition and Understanding Proceedings.

[6]  Richard M. Schwartz,et al.  A compact model for speaker-adaptive training , 1996, Proceeding of Fourth International Conference on Spoken Language Processing. ICSLP '96.