BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Error Approximation and Minimum Bayes Risk Acoustic Model Estimati
 on - Matt Gibson  (University of Sheffield)
DTSTART:20080710T120000Z
DTEND:20080710T130000Z
UID:TALK12840@talks.cam.ac.uk
CONTACT:Dr Marcus Tomalin
DESCRIPTION:Acoustic model estimation using the Bayes risk criterion invol
 ves  \ncomputation of the minimum edit (or Levenshtein) error between the 
  \ncorrect sequence of words (or phonemes) and each member of a set of  \n
 alternative hypotheses. In the context of large vocabulary speech  \nrecog
 nition\, the set of alternative hypotheses is large\, rendering the  \ncom
 putational side of this process prohibitively expensive. So  \npractical i
 mplementations of Bayes risk minimization use two  \napproximations.\n\nTh
 e first approximation is to use a pruned set of alternative  \nhypotheses 
 instead of the full hypothesis space. The second is to use  \na computatio
 nally-inexpensive approximation to the Levenshtein error  \nfor each membe
 r of this pruned set.\n\nThis talk focuses on the latter approximation. A 
 previously introduced  \napproach (D. Povey\, PhD thesis\, 2003) uses alig
 nments of the correct  \nand alternative sequences to approximate the Leve
 nshtein error between  \nthe sequences. We examine the accuracy of this ap
 proach\, highlight  \nsome limitations\, and propose alternative alignment
 -based  \napproximations. We then present experimental work quantifying th
 e  \nimpact of these alternatives upon MBR parameter re-estimation.\n
LOCATION:LR5\, Engineering Department\, Baker Building
END:VEVENT
END:VCALENDAR