On the Use of Prior Knowledge in Normalization Schemes for Speaker Verification

✍ Scribed by Guillaume Gravier; Jamal Kharroubi; Gérard Chollet

• Publisher: Elsevier Science
• Year: 2000
• Tongue: English
• Weight: 164 KB
• Volume: 10
• Category: Article
• ISSN: 1051-2004
• DOI: 10.1006/dspr.1999.0366

✦ Synopsis

Recent research on text-independent speaker verification systems has shown that prior knowledge of some source of variability can be used for normalization in order to improve the performance of the systems. In particular, it has been observed that the problem of the handset microphone mismatch between training and testing data can efficiently be treated that way. In this paper, we review how prior knowledge can be included at different normalization levels. We study the benefit of including knowledge of the target speaker gender and, eventually, of the test segment handset type, with experiments on the Switchboard corpus. The experiments pointed out the fact that score normalization with priors improves the separability between genuine speakers and impostors while the additional use of background models with priors increases the stability of the decision boundary.