<p>Artificial emotion recognition is a highly relevant subdomain of human–computer interaction research. This study proposes a novel approach, termed fine-grained temporal factorized bilinear pooling (FG-T-FBP) fusion, to address key challenges in audio–video emotion recognition. Specifically, the proposed model focuses on two major challenges: (1) the effective extraction of fine-grained, context-aware features from both modalities, and (2) their effective fusion for enhanced emotional understanding. To deal with these challenges, modality-specific deep neural networks are employed for feature extraction. A 2D convolutional neural network combined with a gated recurrent unit (GRU) is used to capture temporal dynamics from audio spectrograms. Concurrently, a ResNet-18 backbone, followed by a GRU, extracts features from video frames and models their temporal evolution. These modules capture both local spatial patterns and fine-grained temporal dynamics within each modality, yielding highly contextualized unimodal representations. The effectiveness of FG-T-FBP is evaluated on two widely used benchmark datasets: Carnegie Mellon University Multimodal Opinion Sentiment Intensity (CMU-MOSI) and Interactive Emotional Dyadic Motion Capture (IEMOCAP). The proposed fusion approach demonstrates improved performance compared to conventional fusion paradigms, including early, late, and hybrid fusion, as well as recent state-of-the-art methods. Experimental results show an accuracy of 83.46% and an F1-score of 82.23% on the IEMOCAP dataset (audio–visual, 4 emotion classes), 78.63% accuracy and an F1-score of 77.34% on IEMOCAP (audio–visual, 6 emotion classes), and 76.31% accuracy on the CMU-MOSI dataset. These results highlight the effectiveness of the proposed fine-grained feature extraction, temporal contextualization, and factorized bilinear fusion in improving emotion recognition performance.</p>