U-NORM Likelihood Normalization in PIN-Based Speaker Verification Systems

Abstract

This paper present a new likelihood normalization technique, entitled U-NORM, for speaker recognition systems based on short utterances. A comparison between this new approach and the widely used Z-NORM is reported and evaluated. Phonetic dependency between the speaker model and the test speech utterances is determined as the main impediment for a good performance of Z-NORM technique. A set of experiments are developed on a specifically acquired PIN-oriented real-users database showing the higher performance of the new technique for PIN based security applications. U-NORM provides a common likelihood scale for all system users allowing speaker independent thresholds that simplify the enrollment process and add robustness to PIN based security applications.