Robustness Study of Free-Text Speaker Identification and Verification

Usable free-text speaker identification and voice verification systems must exhibit robustness under varying operational conditions. We study the degree of robustness provided by various signal processing techniques [1] [2] [3] by experimenting on a widely used long distance telephone data base [4] [5] [6]. This data base consists of data recorded at two different sites, with data from one site much poorer in quality than the other; further, the recording equipment had been inadvertly changed for the later half of the sessions resulting in a significantly changed environment. Our study identifies the combination of techniques that provide consistent and significant improvements; our results surpass other published results [4] [5] [6] on the same task. Specifically, in the task of identifying 16 speakers, with training data from the recording prior to equipment change and testing on data from a set after the change(the most challenging condition), we obtain a correct identification rate of 87.5% with an average rank of 1.12; [4] obtains the hitherto best result of 75% correct identification with an average rank of 1.56: without any robustness processing, the rate was only 12%. Detailed results on exhaustive experimentation are presented along with appropriate discussions.