EXAM2: Extending Audio Understanding in Multilingual and Multimodal Analysis

Recent progress in Large Audio Language Models (LALMs) has been hampered by benchmarks that are primarily English-centric and limited to single audio modalities. The EXAM2 benchmark addresses these gaps by providing a comprehensive, multilingual, and multimodal evaluation framework. It spans six languages (English, German, Spanish, Japanese, Malay, and Chinese) and incorporates diverse audio inputs—including speech, sound, music, and mixed-audio settings—paired with visual representations to enable realistic scene-aware reasoning.

EXAM2 consists of 5,667 multiple-choice questions (MCQs), 22,614 image instances, and 135,684 multilingual translations. The construction pipeline emphasizes ecological validity by sourcing real-world recordings and employing rigorous human validation for both textual translations and generated visual counterparts. This structure allows researchers to evaluate how models navigate the intersection of linguistic context, acoustic signals, and visual grounding.

Experimental results reveal a significant performance gap in multilingual and cross-modal understanding among current state-of-the-art models. The study demonstrates that visual grounding provides complementary semantic cues that help disambiguate complex or ambiguous audio events, particularly in environmental sound tasks. Furthermore, the authors introduce Gemma3n-EXAM2, a lightweight fusion model fine-tuned using their proposed OmniLoRA approach. This model achieves substantial gains, including up to 12.4% improvement in multilingual settings and 21.7% in multimodal evaluation, proving that multilingual training acts as a powerful regularizer that enhances performance even in dominant languages like English.

EXAM2 highlights that current audio intelligence is heavily biased toward high-resource Western language ecosystems, with notable performance degradation in East Asian languages. By establishing a unified, multilingual, and multimodal benchmark, this work provides a critical tool for developing more equitable and robust audio-understanding systems that can generalize across diverse cultural and acoustic environments.