Comparing Emotional Signatures in Tamil Indie and Mainstream Songs Using Neuro Symbolic AI

Abstract

This study introduces a neurosymbolic AI approach to examining emotional expression in Tamil romance songs, addressing the broader gap in how non-Western musical emotions are represented in computational models. By combining machine learning-based audio pattern recognition with symbolic reasoning rules generated via large language models, the framework bridges statistical learning and cultural interpretation. The approach is applied to 117 independently produced and mainstream commercial Tamil romance tracks. Audio features such as tempo, energy and spectral centroid are extracted and interpreted using emotion classification rules generated by GPT-4o-mini, whose outputs are treated as LLM-consistent symbolic labels for exploratory analysis. These rules serve as a basis for symbolic inference, enabling the system to logically deduce the emotion conveyed by a song segment based on its audio attributes. Three new tools are introduced: the Emotional Authenticity Index (EAI), assessing the richness of emotional expression by aggregating entropy-based diversity and symbolic variety; the Symbolic Logic Consistency (SLC), evaluating the coherence of symbolic audio signatures; and Temporal Emotion Flow (TEF) visualisations, mapping emotional evolution across a song. Dimensionality reduction using UMAP on symbolic emotion scores reveals distinct emotional clustering by emotion with no meaningful separation by genre. Emotion categories align with the Geneva Emotional Music Scale (GEMS), a validated framework that captures music-specific emotions such as nostalgia, transcendence and tenderness. A Classification and Regression Tree (CART) decision tree model is used to classify symbolic emotion categories based on three key audio features, revealing which attributes most strongly influence the expression of specific emotions. Preliminary findings suggest that while Tamil indie and mainstream music may differ in symbolic strategies, both appear to exhibit convergent emotional trajectories consistent with a shared cultural grammar of emotion. These insights not only enrich cross-cultural musicology but also demonstrate how interpretable AI may illuminate the cultural logic of affective expression, expanding the scope of emotion-aware computational modelling.