Comparison of Multiple Classifiers for Audio-Visual Speaker Recognition System

AV-ASR (Audio-Visual Automatic Speaker Recognition) is an evolving field of study in person authentication. Relying solely on audio or video features for recognition is susceptible to interference from background noise. Therefore, the integration of both modalities is necessary to enhance accuracy. In this paper, we introduce a novel set of hybrid visual features based on various information presented in the form of appearance and shape features. These visual features fall into three primary categories: texture-based, appearance-based, and shape-based features. For appearance-based visual features, DCT (Discrete Cosine Transform) is employed and for shape-based features, AAM (Active Appearance Model) are used. The experiments in AV-ASR employ a classification approach, where the results of various classifiers are assessed and compared. Among all the classifiers used in this paper, we achieve the highest recognition accuracy of 93.73% using the proposed AV-ASR model.