Speaker verification from actual telephone voice

Abstract

One of the most convenient and direct ways to identify people is talker verification. This is virtually the only method feasible for telephone handset user identification without resorting to additional input devices. One of the remaining unsolved obstacles to the practical use of such a verification technique is the lack of a method to reduce the effects of variations in transmission characteristics due to changes introduced by the telephone connection path or handset utilized. A new means of average self inverse filtering to normalize the effects of linear distortion of actual telephone voice is proposed in this paper. The combined use of average self inverse filtering and selected band analysis improves verification performance beyond that obtained with PARCOR parameters directly extracted from actual telephone voice. Better than 95% verification is possible by the additional use of pitch frequency. In addition, if only one type of telephone handset is utilized almost 100% verification is attainable.