Boosting speaker identification performance using a frame level based algorithm

Abstract

In this paper, we propose an algorithm to improve the performance of speaker identification systems. A baseline speaker identification system uses a scoring of a test utterance against all speakers' models; this could be termed as an evaluation at the observation level. In the proposed approach, and prior to the standard evaluation phase, an algorithm based on a frame level evaluation is applied. The speaker identification study is conducted using IVIE corpus and a randomly selected 120 speakers from TIMIT. Mel-frequency cepstral coefficients (MFCC) and Gaussian mixture model (GMM) are the main components in state of the art speaker identification systems and will be adopted in this work. Experimental results based on several systems with different training and testing conditions, showed that our proposed algorithm yielded to relative reduction in error rates of 24.4 and 37.3% over the baseline systems respectively for IVIE and TIMIT. The final performances reached measured by identification error rates are 3.4% and 5.2% for IVIE and TIMIT corpuses.