Presentation 1994/1/21
On-line Speaker Adaptation using Maximum a posteriori Estimation
Tatsuo Matsuoka, Chin-Hui Lee,
PDF Download Page PDF download Page Link
Abstract(in Japanese) (See Japanese page)
Abstract(in English) In this paper,we study issues related to the use of on-line incremental adaptation for speech recognition.We apply the MAP estimation that uses conjugate prior as prior distribution.That conjugate prior is known as sufficient statistics for exponential family.We use the segmental MAP algorithm to perform HMM adaptation.Two modes of incremental adaptation,namely supervised and unsupervised adaptation,are tested.In the supervised mode,the correct word transcription is used for MAP adaptation,while in the unsupervised mode,the word transcription is provided by the recognizer.We report on the results obtained for the ARPA Air Travel Information stem(ATIS)task.Compared with the speaker independent recognition results,the supervised and unsupervised incremental adaptation algorithms reduce the word error rate by 25% and 5.6% respectively.
Keyword(in Japanese) (See Japanese page)
Keyword(in English) speaker adaptation / hidden Markov model(HMM) / maximum aposteriori estimation(MAP)
Paper # SP93-133
Date of Issue

Conference Information
Committee SP
Conference Date 1994/1/21(1days)
Place (in Japanese) (See Japanese page)
Place (in English)
Topics (in Japanese) (See Japanese page)
Topics (in English)
Chair
Vice Chair
Secretary
Assistant

Paper Information
Registration To Speech (SP)
Language JPN
Title (in Japanese) (See Japanese page)
Sub Title (in Japanese) (See Japanese page)
Title (in English) On-line Speaker Adaptation using Maximum a posteriori Estimation
Sub Title (in English)
Keyword(1) speaker adaptation
Keyword(2) hidden Markov model(HMM)
Keyword(3) maximum aposteriori estimation(MAP)
1st Author's Name Tatsuo Matsuoka
1st Author's Affiliation NTT Laboratories()
2nd Author's Name Chin-Hui Lee
2nd Author's Affiliation AT&T Bell Laboratories
Date 1994/1/21
Paper # SP93-133
Volume (vol) vol.93
Number (no) 427
Page pp.pp.-
#Pages 8
Date of Issue