Skip to main content
EURASIP Journal on Audio, Speech, and Music Processing

Table 2 Comparison of cosine scoring with different compensation and normalization techniques

From: An improved i-vector extraction algorithm for speaker verification

 

Male

 

English trials

All trials

 

EER(%)

DCF

EER(%)

DCF

LDA(210)

2.50

0.0246

5.57

0.0538

LDA(210), R = 1000

2.48

0.0245

5.48

0.0533

LDA(210), R = 975

2.31

0.0226

5.37

0.0521

LDA(210), R = 950

2.27

0.0220

5.24

0.0514

LDA(210), R = 925

2.23

0.0211

5.18

0.0510

LDA(210), R = 900

2.15

0.0207

5.10

0.0503

LDA(210), R = 875

2.30

0.0224

5.12

0.054

LDA(210), R = 850

2.53

0.0250

5.17

0.510

EFRnorm, LDA(210)

2.35

0.0233

5.40

0.0525

EFRnorm, LDA(210), R = 900

2.29

0.0225

5.19

0.0512

  1. All results are obtained with cosine scoring, LDA dimension of 210, and without EFR normalization. The baselines of our systems are 2.35 % for EER in English trials and 5.40 % for EER in all trials (italics), where EFR normalization and LDA = 210 are applied, best results after applying CRA are 2.15 % for EER in English trials and 5.10 % for EER in all trials (italics), where only LDA = 210 are applied
Back to article page