Robust features for text-independent speaker recognition with short utterances

Abstract

Speaker recognition systems achieve good performance under controlled conditions. However, in real-world conditions, the performance degrades drastically. The principal cause being when limited data are presented. The presence of background noise is another main factor of performance distortion. In spite of the major advances in speaker recognition field, the effect of noise and the limitation of the amount of available speech data are still open problems, and no optimal solution has been found yet to cope with them. In this paper, we propose a new system using new enhanced and reduced gammatone coefficients in order to improve robustness with limited speech data duration. We demonstrate the usefulness of these coefficients compared to the well-known features with speakers taken from different databases recorded under different conditions.