BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Unsupervised Learning from Users' Error Correction in Speech Dicta
 tion - 
DTSTART:20060518T100000Z
DTEND:20060518T110000Z
UID:TALK5024@talks.cam.ac.uk
CONTACT:Oliver Stegle
DESCRIPTION:http://www.cs.indiana.edu/~doyu/listenrain/LearnFromCorection-
 ICSLP2004.pdf\n\nAbstract:\nWe propose an approach to adapting automatic s
 peech recognition systems used\nin dictation systems through unsupervised 
 learning from users' error\ncorrection. Three steps are involved in the ad
 aptation: 1) infer whether the\nuser is correcting a speech recognition er
 ror or simply editing the text\, 2)\ninfer what the most possible cause of
  the error is\, and 3) adapt the system\naccordingly. To adapt the system 
 effectively\, we introduce an enhanced\ntwo-pass pronunciation learning al
 gorithm that utilizes the output from both\nan n-gram phoneme recognizer a
 nd a Letter-to-Sound component. Our\nexperiments show that we can obtain g
 reater than 10% relative word error\nrate reduction using the approaches w
 e proposed. Learning new words gives\nthe largest performance gain while a
 dapting pronunciations and using a cache\nlanguage model also produce a sm
 all gain.
LOCATION:Room 911\, Rutherford Building\, Cavendish Laboratory\, Departmen
 t of Physics
END:VEVENT
END:VCALENDAR