| Abstract: | We present our ongoing research focused on speaker recognition in historical oral archives. This research is part of our long-term effort aimed at enabling versatile access to the archive of the Czech Radio (CRo). Based on a manually annotated partition of the archive, we compiled a database covering a time span of more than 30 years to carry out our experimental study. Hence we were able to investigate the impact of various aspects that make it challenging to process historical data. We show the shift of scores for target (genuine) speaker trials introduced by the aging effect, the value of the signal-to-noise ratio or by the variable amount of the enrollment and test data. Scores for speaker detection trials were assessed by a system based on the i-vector paradigm and probabilistic linear discriminative analysis. We also assessed the performance of this system using an evaluation database containing contemporary recordings collected over a time span of approximately 4 years. Although using state-of-the-art techniques, capable of dealing with nuisance inter-session variability, we demonstrate remarkable degradation in the performance of the system in the evaluation containing historical data compared to the one containing contemporary data only. Specifically, the Equal Error Rate (EER) of the system rose to 8.27 % from 1.93 %. The revealed difference thus exemplifies that compensation techniques need to be employed to cope with additional variability introduced in the historical data by various sources. |
