ISCA Archive Odyssey 2001
ISCA Archive Odyssey 2001

Noise robustness in forensic speaker verification

Yosef A. Solewicz

Recently, commercial speaker verification systems have been applied to forensic casework. Unfortunately, these systems were not custom tailored to current needs. In this paper, we try to identify general guidelines towards establishing an improved forensic speaker verification methodology. A main issue is the recognition rate, which highly deteriorates in noisy conditions, preventing a reliable correspondence between LLR values and a desirable significance scale. It is shown that segmented classification in time and frequency efficiently decrease noise related LLR variance, thus improving performance in forensic applications.