Detection of community emotions through Sound: An Investigation using the FF-Orbital Chaos-Based feature extraction model

Abstract

Community Emotion Detection (CED) is important to public safety and sound forensics. It enables the analysis of collective emotions in dynamic environments such as public events, protests, and emergencies. This research presents a new, cost-effective, automatic, and self-organized classification model for CED. The major objective of this research is to show the effectiveness of signal processing in the CED research area. The largest known CED sound dataset was constructed for this study. This dataset was curated to investigate the presented automated CED model's classification capability. The curated CED sound dataset consists of 5,051 three-second overlapping sound samples categorized as negative, neutral, and positive emotions. This dataset increases the model's generalizability by reflecting diverse acoustic and emotional scenarios. A self-organized feature extraction function, called the Forward-Forward Pattern-Based Feature Generator (FFOrbital), is introduced. The FF-Orbital autonomously selects the most suitable pattern from six predefined patterns. This eliminates the need for manual feature engineering. Additionally, a multi-level feature extraction method is enabled by integrating the Unbalanced Tree Multilevel Discrete Wavelet Transform (UTMDWT). This method generates frequency bands that provide extract spatial and frequency-domain features. Iterative Neighborhood Component Analysis (INCA) has selected the most informative features. INCA is a self-organized feature selector. Classification is then performed using a Bayesian-optimized Support Vector Machine (SVM). Tests were conducted using 10-fold cross-validation. The model achieved a classification accuracy of 98.81%. These results demonstrate the usability of the CED model and its effectiveness in digital forensics, public safety, and community-level sentiment analysis. This work makes a significant contribution to the CED research domain by providing a feature-engineeringbased alternative to resource-intensive deep learning models. The results show that the proposed model is valuable for signal processing and sound forensics.