 |
ITRW on
|
 |
|
ITRW on
|
|
This paper presents the second part of a new approach to on-line unsupervised adaptation in speaker verification. The new approach extends previous work in the literature by (1) improving performance on the enrollment handset-type when adapting on a different handset-type (e.g., improving performance on cellular when adapting on a landline office phone), (2) accomplishing this cross channel improvement without increasing the size of the speaker model after adaptation, (3) employing a countbased, parameter-dependent smoothing algorithm that emphasizes the use of mean parameters in the speaker models until sufficient adaptation data are present to accurately estimate variances, and (4) developing a new confidence-based adaptation update weight which minimizes the corrupting effects on the speaker models from impostor attacks. Experimental results show a 61% (rel.) overall reduction in EER using the new on-line adaptation approach even with a significant impostor attack rate, and a 24% improvement in EER due to the new confidencebased adaptation scheme for those speaker models corrupted by impostor utterances.
Bibliographic reference. Heck, Larry / Mirghafori, Nikki (2001): "On-line unsupervised adaptation in speaker verification: Confidence-based updates and improved parameter estimation", In Adaptation-2001, 53-56.