Multimodal Emotion Recognition Using Feature Fusion: An LLM-Based Approach

Multimodal emotion recognition is a developing field that analyzes emotions through various channels, mainly audio, video, and text. However, existing state-of-the-art systems focus on two to three modalities at the most, utilize traditional techniques, fail to consider emotional interplay, lack the scope to add more modalities, and aren’t efficient in predicting emotions accurately. This research proposes a novel approach using rule-based systems to convert non-verbal cues to text, inspired by a limited prior attempt that lacked proper benchmarking. It achieves efficient multimodal emotion recognition by utilizing distilRoBERTa, a large language model fine-tuned with a combined textual representation of audio (such as loudness, spectral flux, MFCCs, pitch stability, and emphasis) and visual features (action units) extracted from videos. This approach is evaluated using the datasets RAVDESS and BAUM-1. It achieves high accuracy (93.18% in RAVDESS and 93.69% in BAUM-1) on both datasets, performing on par with the SOTA (state-of-the-art) systems, if not slightly better. Furthermore, the research highlights the potential for incorporating additional modalities by transforming them into text using rule-based systems and utilizing them to refine further pre-trained large language models, giving rise to a more comprehensive approach to emotion recognition.