Towards measuring uniqueness of human voice

The use of voice as a biometrie modality for user authentication and identification has grown very rapidly. It is therefore very important that we understand limitations of such systems which will ultimately depend on the discriminative power of the voice biometric. In this paper, we have contributed towards measuring distinctiveness of voice biometric by both formulating a new measure and creating a new dataset to perform more reliable measurements. For this purpose, we evaluate the prominent approaches in the field and propose a new approach that better incorporates within-user variability and is analytically more tractable. Our newly created dataset includes voice samples extracted from close to two thousand TED Talks videos. Overall our measurements on this dataset revealed a biometric information content of about 60 bits in human voice. Further, tests performed by adding some generic voice effects on the samples show that the distinctiveness reduces by almost 20 bits, implying that when true variability is reflected in user samples resulting entropy may further reduce.