Enhanced speech emotion understanding using advanced attention-centric convolutional networks

Abstract

Speech Emotion Recognition (SER) plays a crucial role in Human-Computer Interaction (HCI) systems, enabling machines to understand and respond to human emotional states. This paper presents an advanced framework leveraging feature fusion and deep learning architectures for robust SER. The proposed model integrates multi-features extracted using techniques such as MFCC, ZCR, and chroma. These features are augmented with statistical summaries including mean, maximum, and minimum values of MFCCs, enhancing the discriminative power of the input representation. The proposed deep learning architecture, Advanced Attention-Centric Convolutional Networks (AACCN), incorporates a hybrid approach combining Multi-Head Attention (MHA) mechanisms with Convolutional Neural Networks (CNNs). MHA is employed to capture intricate dependencies within the input sequences, while CNNs facilitate hierarchical feature learning and spatial modeling of temporal sequences. Batch normalization and dropout are applied to enhance model generalization and mitigate overfitting. Experimental results on benchmark datasets demonstrate that the proposed framework achieves state-of-the-art performance in SER tasks. Results show significant improvements in accuracy, precision, recall, and F1-score metrics compared to baseline models. The effectiveness of feature fusion and the synergy between MHA and CNNs highlight the robustness and scalability of the proposed AACCN model across diverse emotional contexts in speech signals.