EFFICIENT SCORE NORMALIZATION FOR SPEAKER RECOGNITION

被引：6

作者：

Aronowitz, Hagai ^{[1
]}

Aronowitz, Vanessia ^{[2
]}

机构：

[1] IBM Haifa Res Labs, IL-31905 Haifa, Israel

[2] Hebrew Univ Jerusalem, Dept Stat, Jerusalem, Israel

来源：

2010 IEEE INTERNATIONAL CONFERENCE ON ACOUSTICS, SPEECH, AND SIGNAL PROCESSING | 2010年

关键词：

speaker recognition; speaker identification; score normalization; T-norm; Z-norm;

D O I：

10.1109/ICASSP.2010.5495629

中图分类号：

O42 [声学];

学科分类号：

070206 ; 082403 ;

摘要：

Score normalization is an important component in most speech classification tasks including speaker recognition. State-of-the-art scoring approaches use both T-norm and Z-norm. This paper addresses the following goals: better understanding of existing score normalization methods, reducing the need for explicit score normalization, and improving the computational efficiency of score normalization. In addition, the importance of score normalization for speaker identification is demonstrated, and accuracy is improved considerably using various normalization techniques.

引用

页码：4402 / 4405

页数：4

共 9 条

[1] [Anonymous], IEEE T AUDIO SPEECH
[2] ARONOWITZ H, 2004, P ICSLP
[3] Aronowitz H., 2007, IEEE T AUDIO SPEECH
[4] Aronowitz H., 2005, P INT
[5] ARONOWITZ H, 2007, P INT
[6] Score normalization for text-independent speaker verification systems
Auckenthaler, R
Carey, M
Lloyd-Thomas, H
[J]. DIGITAL SIGNAL PROCESSING, 2000, 10 (1-3) : 42 - 54
[7] Ben M., 2002, P ICASSP
[8] Speaker verification using adapted Gaussian mixture models
Reynolds, DA
Quatieri, TF
Dunn, RB
[J]. DIGITAL SIGNAL PROCESSING, 2000, 10 (1-3) : 19 - 41
[9] Sturim D., 2005, ICASSP

← 1 →