Performance Analysis of CBAM-Based Capsule Net for Neural Network for Speech Emotion Recognition

Emotion recognition from audio signals is a challenging yet crucial task with applications in human-computer interaction, affective computing, and psychological research. This paper presents an innovative approach for audio emotion recognition, starting with a comprehensive pre-processing pipeline that integrates Infinite Impulse Response (IIR) filtering, MFCC, and Modified Emphasized Dynamic Com (MEDC). This combination surpasses conventional methods like MFCC and FFT by better isolating emotion-specific features and reducing noise. The paper introduces a novel architecture combining Capsule Networks (CapsNets) with a Convolutional Block Attention Module (CBAM). The CapsNet architecture, inspired by the human visual system, efficiently captures hierarchical spatial features and contextual dependencies, addressing the limitations of traditional CNN-based models. The integration of CBAM further refines the feature maps by emphasizing salient regions, improving emotion-related information extraction. The proposed system achieves an accuracy of 98.57% in recognizing emotions from audio data. Experimental results demonstrate the effectiveness of this approach on benchmark datasets, showing resilience to variations in voice quality, background noise, and speaker characteristics. A comparative analysis with traditional deep learning architectures and existing emotion recognition methods substantiates the CapsNet-CBAM model’s accuracy and computational efficiency superiority.