Temporal Misalignment in Affective AI: A Time-Aware Approach to Emotion Recognition

The study analyzes the critical problem of temporal misalignment in AI-based emotion recognition systems as the current models show insufficient capability when detecting dynamic emotional transitions. The misclassification rates remain high when Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) analyze static or short-range patterns with littlesuccess during the shifting of emotions. This paper introduces Time-Aware Emotion Recognition Model (T-AERM) as a model which uses Vision Transformers (ViT) for spatial feature retrieval simultaneously with Temporal Convolutional Networks (TCN) to handle long-range sequential dependencies in emotion recognition applications. The proposed attention-based temporal alignment technique enables the model to understand the emotional progression between sequential frames. The T-AERM model surpasses previous models in performance by reaching 31% higher accuracy rates while observing 35% fewer false positives when evaluated on AFEW Affect Net and RAVDESS datasets. Research findings show that unification of spatial and temporal attention methods enables the development of affective computing systems which process context information effectively across healthcare fields and surveillance systems and human-computer interaction domains.