A Correction of Missing Reliability for Robust Bimodal Speaker Identification

摘要

Speaker identification in real environment is a key issue in biometrics technology for human computer interaction. In this paper, we propose a fuzzy membership function for adaptive threshold in different modalities reliability measure for robust bimodal speaker identification. In the bimodal speaker identification system, we will also propose an extension of a modified convection reliability function applied to both the audio and lip information to account optimal reliability simultaneously for audio and visual information integration. For creating mismatch in between train and test data, babble noises and artificial illumination have been added to test speeches and lip images, respectively. Local PCA have been applied at features level to both classifiers system for reducing the dimension of feature vector at different stage of signal distortion. We have applied particle swarm optimization (PSO) for optimizing the proposed fuzzy based adaptive threshold and modified convection function's optimizing parameters. The entire speaker identification experiments have been performed using VidTimit database. Experimental results show that our proposed method enhanced the identification accuracy in comparison with the baseline system thus demonstrated the validation of the proposed approach and most notably maintains the consistency of the integration process.