Overview of Speaker Identification with Privacy

Abstract

Speaker identification is the task of determining the identity of the speaker given a test speech sample. We typically associate the identity of the speaker with a predetermined set of speakers and we refer to this task as closed-set speaker identification. If we augment the set of speakers with a none of the above option, i.e., consider the speaker to be outside the predetermined set of speakers, this task becomes open-set speaker identification. Speaker verification can be considered as a generalization of open-set speaker identification, where the set of speakers is restricted to one speaker. As a consequence, the algorithms used in speaker verification can be extended for speaker identification. The main difference between the two tasks is their application scenarios and this results in different evaluation metrics for the two tasks. Speaker verification is mainly used for authentication; speaker identification finds use in surveillance applications and also as a preliminary step for other speech processing applications as we discuss below.