Distribution of speaker verification errors in time and (speaker) space
This seminar will present results from automatic speaker verification experiments with data from PER, a physical access control system mounted at the entrance to TMH, with special emphasis on the analysis of error rates for individual users. Two Bayesian estimation methods have been used. The first estimates error rates directly from observed decision errors, while the second estimates the mean and variance of a parametric (normal) distribution of observations on the decision variable. Results show how error rates vary with the user and with time from users' enrollment sessions.