Robust speaker verification in low bit rate channels

Abstract

The coder mismatch problem in speaker verification systems is addressed. A speech coder detector is proposed that correctly classifies the coding scheme used in a given utterance having a length of 3 s with an average rate of 93%. It is shown that, with the use of the proposed speech coder detector, the average speaker verification accuracy over a pool of four different coders used in low bit rate communication channels is increased close to the average accuracy obtained in the ideal matched case.