Vector Angle Minimum Criteria for Classi¯erSelection in Speaker Veri¯cation Technology
-
Graphical Abstract
-
Abstract
Decision level fusion between classi¯ers
plays an increasingly important role in present speaker ver-
i¯cation technology, while on the other hand, it is not the
more classi¯ers involved in the fusion, the better the sys-
tem performs. This paper proposes the Vector angle min-
imum (VAM) criteria for the classi¯er selection in score
fusion of speaker veri¯cation system. The main idea is to
study the directions of the score-vectors in score space, ¯nd
the score-vectors which can form a minimum angle with
the standard-vector by linear combination, and take the
corresponding classi¯ers into the fusion. The experimen-
tal results show that the VAM-based selection can reduce
the needed number of classi¯ers obviously and enhance the
system performance. When compared to the n-best crite-
ria selection, the Equal error rate (EER) of the fused sys-
tem is relatively 7.2% lower when the number of selected
classi¯ers is 16.
-
-
DownLoad: