Fuzzy Audio-Visual Feature Maps for Speaker Identification

摘要

Speech-based person recognition by machine has not reached the level of technological maturity required by some of its potential applications. The deficiencies revolve around sub-optimal pre-processing, feature extraction or selection, and classification, particularly under conditions of input data variability. The joint use of audible and visible manifestations of speech aims to alleviate these shortcomings, but the development of effective combination techniques is challenging. This paper proposes and evaluates a combination approach for speaker identification based on fuzzy modelling of acoustic and visual speaker characteristics. The proposed audio-visual model has been evaluated experimentally on a speaker identification task. The results show that the joint model outperforms its isolated components in terms of identification accuracy. In particular, the cross-modal coupling of audio-visual streams is shown to improve identification accuracy.