Robustness Study of Free-Text Speaker Identification and Verification

Abstract

Usable free-text speaker identification and verification systems must exhibit robustness under varying operational conditions. We studied the degree of robustness provided by various signal processing techniques - spectrum subtraction, bandpass liftering, RASTA filtering, ISDCN, and stereo database normalization. The experiments were performed on a widely used, challenging long distance telephone database. This database 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 inadvertently changed for the later half of the sessions resulting in a significantly changed environment. Our study identifies the combination of techniques that provides consistent and significant improvements; our results surpass other published results on the same task. We further verified the results on two other databases and achieved consistent improvements. Detailed results on exhaustive experimentation are presented along with appropriate discussions.