Sixth European Conference on Speech Communication and Technology
(EUROSPEECH'99)

Budapest, Hungary
September 5-9, 1999

Enhanced Likelihood Computation Using Regression

Peter de Souza, Bhuvana Ramabhadran, Yuqing Gao, Michael Picheny

IBM T. J. Watson Research Center Yorktown Heights, NY, USA

In a rank based large vocabulary continuous speech recognition system [1], the correct leaf is expectedto occupy the top rank positions. An increase in thenumber of times the correct leaf occurs in the top rankpositions translates to an increase in word accuracy.In order to achieve low error rates, we need to discriminate the most confusable incorrect leaves fromthe correct leaf by lowering their ranks. Therefore, the goal here is to increase the likelihood of the cor-rect leaf of a frame, while decreasing the likelihoodsof the confusable leaves. In order to do this, we usethe auxiliary information from the prediction of theneighboring frames to augment the likelihood computation of the current frame. We then use the residual errors in the predictions of neighboring frames todiscriminate between the correct (best) and incorrectleaves of a given frame. In this paper, we presenta new algorithm that incorporates prediction errorlikelihoods into the overall likelihood computation toimprove the rank position of the correct leaf. Experimental results on the Wall Street Journal task and anin-house large vocabulary continuous speech recogni-tion task show a relative accuracy improvements inspeaker-independent performance of 10%.


Full Paper (PDF)   Gnu-Zipped Postscript

Bibliographic reference.  Souza, Peter de / Ramabhadran, Bhuvana / Gao, Yuqing / Picheny, Michael (1999): "Enhanced likelihood computation using regression", In EUROSPEECH'99, 1699-1702.