It is now well established from a variety of studies that there is asignificant benefit from combining video and audio data in detecting activespeakers. However, either of the modalities can potentially mislead audiovisualfusion by inducing unreliable or deceptive information. This paper outlinesactive speaker detection as a multi-objective learning problem to leverage bestof each modalities using a novel self-attention, uncertainty-based multimodalfusion scheme. Results obtained show that the proposed multi-objective learningarchitecture outperforms traditional approaches in improving both mAP and AUCscores. We further demonstrate that our fusion strategy surpasses, in activespeaker detection, other modality fusion methods reported in variousdisciplines. We finally show that the proposed method significantly improvesthe state-of-the-art on the AVA-ActiveSpeaker dataset.