Automatic emotion recognition from video streams is an essential challenge for various applications including human behavior understanding, mental disease diagnosis, surveillance, or human-machine interaction. In this paper we introduce a novel, completely automatic, multimodal emotion recognition framework based on audio and visual fusion of information designed to leverage the mutually complementary nature of features while maintaining the modality-distinctive information. Specifically, we integrate the spatial, channel and temporal attention into the visual processing pipeline and the temporal self-attention into the audio branch. Then, a multimodal cross-attention fusion strategy is introduced that effectively exploits the relationship between the audio and video features. The experimental evaluation performed on RAVDESS, a publicly available database, validates the proposed approach with average accuracy scores superior to 87.85%. When compared with the state-of the art methods the proposed framework returns accuracy gains of more than 1.85%.