A multi-sample single-source model using spectrographic features for biometric authentication

In this paper we propose a novel approach by using spectrographic features and correlation filters as classifiers to perform speaker authentication. Visual displays (spectrograms) from speech signals produced from different persons are used as features for the verification task. In order to achieve better verification results, the exclusion of low energies and the inclusion morphological image processing steps are applied to the spectrograms. It is discovered that, by applying these two techniques, the verification performance improves significantly. For the classification modeling, Unconstrained Minimum Average Correlation Energy (UMACE) filter is implemented. We propose a multi-sample approach by fusing multiple samples from different utterances at the score level. By using the average operator, both the theoretical and empirical results show that by integrating as many samples as possible can improve the overall reliability of the system. This model is called as multi-sample single-source (MSSS) model. A digit database has been used for performance evaluation, yielding an overall performance of 99.6%.