Human vs. machine speaker identification with telephone speech
Astrid Schmidt-Nielsen, Thomas H. Crystal · 1998
An experiment compared the speaker recognition performance of human listeners to that of computer algorithms/systems. Listening protocols were developed analogous to procedures used in the algorithm evaluation run by the U.S. National Institute of Standards and Technology (NIST), and the same telephone conversation data were used. For "same number" testing, with three-second samples, listener panels and the best algorithm had the same equal-error rate (EER) of 8%. Listeners were better than typical algorithms. For "different number" testing, EER's increased but humans had a 40% lower equalerror rate. Other observations on human listening performance and robustness to "degradations" were made. 1. INTRODUCTION The development of automatic speaker verification technology has been greatly facilitated by the recent development of uniform training and test materials and of procedures and metrics for evaluating verification performance [1,2]. For many, the performance goal for automatic verif...