Presentation 2001/8/24
The SVD-based MLLR speaker adaption method
Toyotsuna Imai, Kazumasa Yamamoto, Hiroshi Matsumoto,
PDF Download Page PDF download Page Link
Abstract(in Japanese) (See Japanese page)
Abstract(in English) In speaker adaption based on the MLLR framework, this paper proposes a SVD-based liner regression which can be applied to smaller regression classes and hence adapt HMMs to spectral dependent speaker differences. In this method, each row vector of a regression matrix is estimated by minimizing the norm under an effective rank based on a singular value decomposition(SVD) of the normal equation. This regressin transformation can be applied to any size of regression classes without ill-conditiones. This constrained MLLR is combined with a tree-based tying technique. Large vocabulary recognition tests show that the proposed method achieves 1% higher word accuracy than the conventional MLLR.
Keyword(in Japanese) (See Japanese page)
Keyword(in English) MLLR / SVD / regresion class / effective rank / dictation system
Paper # SP2001-50
Date of Issue

Conference Information
Committee SP
Conference Date 2001/8/24(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) The SVD-based MLLR speaker adaption method
Sub Title (in English)
Keyword(1) MLLR
Keyword(2) SVD
Keyword(3) regresion class
Keyword(4) effective rank
Keyword(5) dictation system
1st Author's Name Toyotsuna Imai
1st Author's Affiliation Faculty of Engineering, Shinshu University()
2nd Author's Name Kazumasa Yamamoto
2nd Author's Affiliation Faculty of Engineering, Shinshu University
3rd Author's Name Hiroshi Matsumoto
3rd Author's Affiliation Faculty of Engineering, Shinshu University
Date 2001/8/24
Paper # SP2001-50
Volume (vol) vol.101
Number (no) 271
Page pp.pp.-
#Pages 8
Date of Issue