Feature mapping based on GMM supervector

Author

Guo, Wu ; Dai, Lirong

Author_Institution

iFly Speech Lab., Univ. of Sci. & Technol. of China, Hefei

fYear

2008

fDate

7-9 July 2008

Firstpage

1081

Lastpage

1085

Abstract

The channel or inter-session variability problem is one of the most important factors causing recognition errors in speaker recognition systems. In this paper, we have proposed three methods to estimate the channel supervector in the GMM supervector space to deal with this problem, namely EM clustering, PCA and NAP algorithms. Furthermore, feature mapping is applied to the MFCC after the estimation of channel supervector. The EER of the feature mapping system decreases by 34% relatively over the baseline GMM system in the NIST 2006 SRE core test.

Keywords

Gaussian processes; expectation-maximisation algorithm; feature extraction; principal component analysis; speaker recognition; EM clustering; GMM supervector space; NAP algorithm; PCA algorithm; channel variability problem; feature mapping; intersession variability problem; recognition error; speaker recognition system; Clustering algorithms; Covariance matrix; Eigenvalues and eigenfunctions; Equations; NIST; Principal component analysis; Space technology; Speaker recognition; Speech; Telephone sets;

fLanguage

English

Publisher

ieee

Conference_Titel

Audio, Language and Image Processing, 2008. ICALIP 2008. International Conference on

Conference_Location

Shanghai

Print_ISBN

978-1-4244-1723-0

Electronic_ISBN

978-1-4244-1724-7

Type

conf

DOI

10.1109/ICALIP.2008.4589964

Filename

4589964