Enhanced fusion methods for speaker verification
Yosef Solewicz and Moshe Koppel
In: 9th International Conference Speech and Computer, 20-22 Sep 2004, Saint-Petersburg, Russia.
This paper presents meta-learning schemes aimed at
improving fusion of low and high level information for
speaker verification in clean and noisy environments.
While traditional systems fuse several classifier outputs
in a uniform fashion independently of test quality, the
proposed schemes use selective fusion weights
according to test quality. A decrease of more than 20%
under noisy conditions and 10% under clean conditions
could be obtained with little calibration.