Analysis of Short-Term Ageing Effects in Speaker Recognition

Finnian Kelly

Tuesday, June 23, 2015
12:30 p.m., Conference Room 5A

The influence of speaker ageing on automatic speaker recognition is an important issue, yet it remains relatively unexplored. This presentation concerns the effect of short-term ageing, over a period of 3 years, on the performance of a current speaker recognition system. A decrease in recognition performance over this period is observed, motivating an approach to score calibration that includes ageing information. Further analysis shows that inter-speaker performance variability increases across ageing. An ‘ageing Zoo’ classification is introduced, allowing the analysis of ageing score trajectories on an individual-speaker level. Finally, some initial links between vocal features and recognition performance across ageing will be highlighted.

Bio:

Finnian Kelly is a research associate at the Center for Robust Speech Systems (CRSS) in The University of Texas at Dallas. Before joining CRSS, he was with the Sigmedia Research group at Trinity College Dublin, where he completed his PhD, ‘automatic recognition of ageing speakers’, in 2014. His main research interests are: speaker recognition in the presence of speaker variability, and the application of speaker recognition to forensics.