Human perception of emotion is based on facial expressions, and speech. Traditional emotion recognition techniques extract features from images or frames (in the case of a video). Our technique, based on the RAVDESS database, creates a hybrid model using features extracted from video, as well as audio provided in the dataset where emotions are being acted out/sung by various actors. We’ve established baseline accuracy with machine learning based approaches for only audio, only video, and audio-video combined features. Our hybrid fusion model, which outperforms all machine learning based approaches, involves pre-trained ResNeXt-101 model-based image feature extraction, and MFCC based audio feature extraction, followed by hybrid fusion architecture for feature combination and output. The dataset is available at -https://zenodo.org/record/1188976#.X62bLGgzaUk
