講演抄録/キーワード |
講演名 |
2011-12-20 10:30
Speaker Verification Using MMAP Adaptation ○Sangeeta Biswas・Johan Rohdin・Koichi Shinoda・Sadaoki Furui(Tokyo Inst. of Tech.) NLC2011-48 SP2011-93 |
抄録 |
(和) |
(まだ登録されていません) |
(英) |
This paper proposes maximum a posteriori (MAP) adaptation of Gaussian mixture models (GMM) using multiple priors for text-independent speaker verification. Although the hierarchical prior used in structural MAP (SMAP) adaptation has been proven to outperform the prior used in relevance MAP adaptation, there may still be some complementary information in the relevance prior which might be useful. An idea of combining these two priors is introduced here in the MAP framework. We call this method multiprior MAP (MMAP). We evaluated our proposed method on NIST SRE 2006 10sec4w-10sec4w task. We compared MMAP with classical maximum likelihood (ML) estimation, relevance MAP and SMAP adaptation techniques and proved its effectiveness. We also investigated the effect of Z-Norm and T-Norm on the likelihood ratio scores of different systems here. |
キーワード |
(和) |
/ / / / / / / |
(英) |
Speaker verification / speaker adaptation / MAP / hierarchical prior / / / / |
文献情報 |
信学技報, vol. 111, no. 365, SP2011-93, pp. 133-137, 2011年12月. |
資料番号 |
SP2011-93 |
発行日 |
2011-12-12 (NLC, SP) |
ISSN |
Print edition: ISSN 0913-5685 Online edition: ISSN 2432-6380 |
著作権に ついて |
技術研究報告に掲載された論文の著作権は電子情報通信学会に帰属します.(許諾番号:10GA0019/12GB0052/13GB0056/17GB0034/18GB0034) |
PDFダウンロード |
NLC2011-48 SP2011-93 |
|