OmniVox demonstrates that general-purpose language models can effectively recognize emotions from speech without specific training, opening new possibilities for security and behavioral monitoring applications.

• Zero-shot capabilities of omni-LLMs match or exceed fine-tuned audio models on emotion recognition benchmarks
• Successfully evaluated across two major datasets: IEMOCAP and MELD
• First systematic evaluation of four different omni-LLMs on speech emotion recognition
• Provides a foundation for advanced security applications in detecting emotional states that could indicate threats or unusual behavior

This research highlights how multimodal AI systems can transform security monitoring by accurately interpreting emotional cues in speech, potentially enabling earlier threat detection and more nuanced behavioral analysis.

Original Paper: OmniVox: Zero-Shot Emotion Recognition with Omni-LLMs