A Statistical Approach to Performance Evaluation of Speaker Recognition Systems

In speaker recognition applications, speaker identification is the process of automatic recognizing who is speaking based on statistical information obtained from speech signals. Considering the limited number of tests in real situations during the classification phase, it is more useful to have an estimator of the probability of error for speaker recognition systems. In this work, we propose a method based on the log-likelihood of each speaker to estimate the probability of error of a speaker recognition system. We assess the performance of the estimator with experimental trials and compare with the actual number of errors. The results show that the performance of our estimator is comparable to the conventional method. The proposed method presents better reliability and fast convergence compared to the counting case. Indeed, we attain an analytical expression for the probability of error that can be used as a gradient for other optimization methods in speaker recognition applications.