Recognizing emotions automatically is important for computers to interact better with people, making interactions feel more natural. However, accurately understanding emotions from videos is challenging. Most of the current methods for recognizing emotions from video depends on older techniques such as 2D-CNNs and RNNs. These techniques are useful; however, they are not able to capture all emotional features. For example, 2D-CNNs are good in extracting spatial details but it fails to detect the movement; conversely, RNNs are better with sequences of data, but it could not extract most of the features from each individual frame. To address these problems, this study proposed a new 3D convolutional neural network (3D-CNNs) for emotion recognition. Unlike the previous mentioned models, 3D-CNNs can extract features from both time and space simultaneously. It helps them to interpret the complex emotional patterns from video frames in a better way. The RAVDESS (Livingstone and Russo in PLoS ONE 13(5):e0196391, 2018) [1] dataset was utilized to evaluate the new approach, which resulted in considerable improvements and an accuracy of 76.04%. This demonstrates how well 3D-CNNs can recognize emotions from videos, showcasing their ability to improve the emotion detection systems. The proposed approach outperformed human accuracy by 1.04% and improved by 4.04% compared to the current state-of-the-art.