The Effects of Several Transmission Systems on an Automatic Speaker Verification System

The automatic speaker verification problem has two aspects--the creation of a reference pattern and the determination of similarity between a test and a reference pattern. When verification is performed over dialed-up telephone lines, the transmission system used in the telephone plant is an additional factor that must be considered. In a recent subjective experiment, 1 the effects of adaptive differential pulse code modulation (ADPCM) coding and linear predictive vocoding (LPC) on the speaker verification accuracy of human listeners was investigated. It was shown that the verification task was easiest (most accurate) when homogeneous systems were used (i.e., the test and 2071 reference utterances were transmitted over the same system) and significantly more difficult (higher error rate) when mixed systems (i.e., different transmission systems) were used for the test and reference utterances. In this paper, we investigate how these same transmission systems affect machine verification accuracy using a system that has been studied for the past few years at Bell Laboratories. 2 " 6 The automatic system is based on the analysis of fixed sentence-long utterances in which the verification features are the time variations (contours) of the pitch and gain (intensity) of the utterance. A training set of utterances (both customer and impostor) is required to establish a reference pattern and to choose weights and measurements for the verification process. Following time alignment of the reference and test contours, a combination of weighted Euclidian distances between a set of test and reference measurements is compared with a threshold to determine whether to accept or reject an identity claim.