ODYSSEY 2004 - The Speaker and Language Recognition Workshop
May 31 - June 3, 2004
The use of a priori speaker-dependent thresholds has been shown convenient for speaker verification. However, their estimation is highly affected by the difficulty of obtaining data from impostors, the mismatched conditions, the scarcity of data in real applications and the need of setting the threshold a priori, during enrollment. In this context, possible outliers, i.e., those client scores which are distant with respect to mean in terms of Log-Likelihood Ratio (LLR), could lead to wrong estimations of client mean and variance. To overcome this problem, we propose here several methods based on pruning LLR scores with different statistical criteria. Before estimating the threshold, score pruning removes outliers and improves subsequent estimations. To solve the problem of impostor data, we also suggest a speaker dependent threshold estimation with only data from clients. Text-dependent and textindependent experiments have been carried out by using a telephonic multisession database in Spanish with 184 speakers, that has been recorded by the authors.
Bibliographic reference. Saeta, Javier R. / Hernando, Javier (2004): "On the use of score pruning in speaker verification for speaker dependent threshold estimation", In ODYS-2004, 215-218.