Adaptive weighting parameter in audio-visual voice activity detection
Matar Buchbinder, Yaakov Buchris, Israel Cohen · 2016
Audio-visual voice activity detectors are traditionally based on fixed algorithms and do not consider the quality of the signals in each modality. This could significantly decrease the detector's performance in cases when one of the signals is relatively of poor quality. We proposed an improved solution, which evaluates the signal's quality in each modality and weights them accordingly. In this paper, we present a method for estimating the video quality, particularly in the presence of noisy motion vectors or global motion of the camera. The fussy motion vectors are intended to simulate blurred, unfocused video or low resolution sensor. An adaptive setting of the weighting parameter between the audio and the video signals ensures an optimal bimodal detector. The proposed method was incorporated with an audio-visual voice activity detector, and was tested with a real data set. Simulation results have shown an improved performance compared to the existing fixed method.