Receiver Operating Characteristic (ROC) Curve & Equal-Error Rate (EER) & Detection Error Trade-Off (DET) Curve & Detection Cost Function (DCF)

⊙ Receiver Operating Characteristic (ROC) Curve & Equal-Error Rate (EER) & Detection Error Trade-Off (DET) Curve & Detection Cost Function (DCF)

 

- Confusion matrix

Predicted/Actual class	Yes	No
Yes	TP	FP
No	FN	TN

 TP (True Positive) : Acceptance of the test speaker who is the same with the target speaker.

FP (False Positive) : Acceptance of the test speaker who is not the same with the target speaker.

FN (False Negative) : Rejection of the test speaker who is the same with the target speaker.

TN (True Negative) : Rejection of the test speaker who is not the same with the target speaker.

→ Utterances of an test speaker result in predicted class.

 

- False acceptance rate (FAR; Miss probability)

The number of verified identities for which the test speaker was different from the target speaker normalized against the total number of acceptances.

In biometrics the instance of a security system incorrectly verifying or identifying an unauthorized person. Also referred to as a type Ⅱ error, a false acceptance typically is considered the most serious of biometric security errors as it gives unauthorized users access to systems that expressly are trying to keep them out.

 

- False rejection rate (FRR; False alarm probability)

The number of identities which were not verified for which the test speaker was the same as the target speaker normalized against the total number of rejections.

 

- Receiver operating characteristic (ROC) curve

In the speaker recognition circles, ROC curve are normally plotted with the FAR on the abscissa and the FRR on the ordinate. In this kind of ROC, the objective is to have the area under the curve approach 0. Generally, a random guess should produce an area of about 0.5.

 

- Equal-error rate (EER)

The percentage where FRR and FAR are same.

 

- Detection error trade-off (DET) curve

The DET curve has been favored over traditional ROC curves since its introduction by NIST in 1997. The DET curve plots the miss probability in percentage form versus the false alarm probability, also in percentages. The DET curve always uses a logarithmic scale.

 

- Detection cost function (DCF)

NIST defines a DCF, , as

where,  is the cost of missed detection,  is the cost of a false alarm, P(Target) is the a-priori probability of a target speaker, and P(NonTarget) is the a-priori probability of a non-target speaker. NIST uses the following values for the 1999 NIST evaluations,

 

                                                                                                                                ※ 참고 서적 : Homayoon Geigi, 2011. Fundamentals of Speaker Recognition. Springer Science+Business Media, LLC, New York.

※ 참고 사이트 : http://www.webopedia.com/TERM/F/false_acceptance.html

※ 참고 사이트 : http://poseidon.csd.auth.gr/LAB_RESEARCH/Latest/SecurityBiometrics.htm

※ 참고 사이트 : http://stackoverflow.com/questions/18436034/plot-a-det-curve