Allowing good impostors to test

Abstract

Biometric testing should attempt to report unbiased, real-world system performance, especially when tested on limited databases. Though testing on a standard database, such as the Linguistic Data Consortiums's YOHO, allows comparison of speaker ver@cation systems, it is well known that certain procedures bias the results low. One such procedure concerns the use of cohort or reference speakers to pe'form veriJication, where the cohort speakers are removed as candidate impostors. A method of testing is proposed to remove this bias by modifying the cohort set for each false acceptance test. Results statistically differ for this mod$ed approach, which tries to "best" model the general population with a fixed random sample. Lastly, three techniques to bound the biometric perjbrmance, using both parametric and noli-parametric resampling is demonstrated.

How to Deal with Multiple-Targets in Speaker Identification Systems?

“…2 one can see that the stack size has a major influence on the group detector performances. For an equal error rate (EER) of 5% we get EER of 12.8% for stack size of 10 and EER of 23.9% for a stack size of 100. In many applications, these error rates are too large to work with; especially if we have very large stack sizes (over 100).…”

Use of Identification Trial Statistics for the Combination of Biometric Matchers

“…The cohort normalizing method [21], [22] considers a subset of enrolled people close to the current test person in order to normalize the score for that person by a log-likelihood ratio of the genuine (current person) and impostor (cohort) score density models. Auckenthaler et al [17] separated cohort normalization methods into cohorts found during training (constrained) and cohorts dynamically formed during testing (unconstrained cohorts).…”

Comparison of combination methods utilizing T-normalization and second best score model

“…The concept of background model has been previously introduced in the speaker identification applications [13,4]. The idea of background models is similar to the idea of identification models -the model should reflect the characteristics of a template with respect to other templates.…”

“…Though earlier developed background models in speaker identification research might include both enrolled and test template models, it is rather convenient to separate them in our research. One example of previous use of background models is the cohort based score normalization for speaker verification [13,4] and for fingerprint verification [2]. Cohort methods find a cohort -a subset of enrolled templates close to the one under consideration as shown by a circle in Figure 1.…”