Multimodal fusion systems combine information from multiple modalities (like audio and image) to improve the accuracy of machine learning models. This work presents a novel multimodal fusion approach for facial emotion recognition from videos, leveraging VGG19 and a compact convolutional transformer for visual feature extraction, combined with a deep convolutional neural network for comprehensive acoustic feature analysis. By employing decision-level soft fusion of probabilistic outputs from individual models trained on independent datasets, the approach improves robustness and accuracy, enabling applications in mental health, customer service, and intelligent tutoring systems. Final testing on a curated subset of the RAVDESS dataset—leveraging both facial expressions and acoustic signals for video-based emotion recognition—demonstrates that the proposed deep multimodal fusion model delivers a clear advantage over single-modality systems. Outperforming unimodal baselines, the model achieves 69.41% accuracy, along with precision, recall, and F1-score values of 0.748, 0.694, and 0.691 respectively, for six distinct human emotions.