A novel approach using deep belief network patterns and attention binary decomposition for automated community emotion detection

Abstract

Context: Sound-based community emotion detection (SCED) estimates community emotion from environmental sounds. It has value for public safety and human-computer interaction. Current SCED models have limited adaptivity on complex audio and often need manual tuning. Objective: We aim to design an accurate and efficient automated SCED model for large-scale data. Methods: We propose a feature extraction framework that combines DBNPat feature generation with ATT-BP attention-driven binary compression. The framework adapts to signal characteristics with low computational cost. We also introduce a new dataset of 10,017 environmental sound clips (three seconds) with negative (n = 1,729), neutral (n = 6,154), and positive (n = 2,134) classes. Results: The proposed SCED model achieves 87.28% accuracy on three-class SCED. It yields 81.30% UAR, 84.71% precision, 82.97% F1, and 80.59% geometric mean on the imbalanced dataset. Conclusion: The model links classical feature design and deep pattern generation in one adaptive pipeline. It offers a practical solution for digital sound forensics and other ambient-audio systems that need fine emotion cues.