We describe a text-independent speaker verification system !MESS that was designed to be highly modular and efficient. The system incorporates an extension of the Tcl script language that facilitates rapid prototyping and high overall execution speed. We provide an overview of the system and its implementation, detailing its applicability as a flexible research environment. Using the system we investigate the importance of parts of the modulation spectrum for speaker verification. We find that higher modulation frequencies are not particularly important for verification -- removing them improves verification performance. On the 1997 Switchboard evaluation set, for test segments of 30 seconds, a relatively small system with lowpass filtering at 10Hz cutoff gives a 12% reduction in EER. 1. INTRODUCTION We are interested in learning what acoustic speech features best characterize a speaker. In particular we are exploring the temporal domain. Unfortunately, because temporal processing re...

Abstract not found