Event

Speaker Recognition System Analysis

Lara Stoll

ICSI

Tuesday, December 02, 2008
12:30

My research focuses on the relation of intrinsic speaker characteristics to automatic speaker recognition system results, with the aim of determining what makes a speaker easier or harder for the system to recognize. For this talk, I will describe my recent work, which has taken two directions. First, using Switchboard-1 data and a basic GMM-UBM system, I investigate the effects of speaker sex, age, dialect area, and education level on score distributions. Secondly, using 2008 Speaker Recognition Evaluation data and the scores from state-of-the-art systems, I look into the phenomenon that target score variance is greater than impostor score variance, and continue with further related experiments.