Speech Emotion Recognition Using Transformer-Based Architectures with Self-Attention Mechanisms
Abstract
Speech Emotion Recognition (SER) has become a key aspect in human-computer interaction, and affective computing, yet, current methods are faced with the challenge of modeling long-range context and fine-grain emotional expressions in speech signals. This paper has countered these shortcomings, giving a Transformer-based framework that incorporates a better self-attention to classify emotions better. The aim is to identify a solid and scalable SER architecture that can help detect various emotional states based on speech data. The procedure consists of combining multimodal acoustic feature extraction (MFCCs, Mel-spectrograms, and prosodic features) with Transformer-based modeling plus adaptive attention enhancement and contrastive learning techniques. Experiments on the Ryerson Audio-Visual Database of Emotional Speech and Song (RAVDESS) show better results, with a goal of 95.1% of accuracy and 94.8 of F1-score, surpassing a number of models at the state-of-the-art. These findings affirm that self-attention based architectures are highly effective at representing emotional features and classifying them. The paper concludes that the suggested strategy is a feasible and efficient method to utilize in any real-world SER application, and the improvements of the approach could be achieved by introducing multimodal and lightweight models.