LLM-Based Pathos Analysis Outperforms Acoustic Emotion Recognition in Political Speech

A study from arXiv (2605.22732) compares acoustic emotion recognition models with large language models (LLMs) for analyzing Pathos in political speech. Using a Bundestag speech by Felix Banaszak (51 segments, 245 seconds), researchers tested three modalities: emotion2vec_plus_large (acoustic SER with Russell Circumplex projection), Gemini 2.5 Flash (LLM analyzing audio and transcript), and TRUST-Pathos (multi-agent LLM supervisor ensemble). Spearman correlations show Gemini Valence strongly correlates with TRUST-Pathos (rho=+0.664, p<0.001), while emotion2vec Valence does not (rho=+0.097, p=0.499). The TRUST pipeline uses three LLM advocates for pathos scoring. Findings suggest LLMs are more effective proxies for rhetorical pathos than traditional acoustic models.