Multi-Domain Audio Question Answering Benchmark toward Acoustic Content Reasoning

We introduce a multi-domain Audio Question Answering (MD-Audio) benchmark designed to advance research in multi-domain sound understanding. The task is organized into three subsets (Bioacoustics, Temporal Soundscapes, and Complex QA), each targeting distinct aspects of interactive question answering over diverse acoustic environments. The benchmark encompasses a broad range of auditory material, including marine mammal vocalizations, environmental soundscapes, and complex real-world audio recordings. We further detail the evaluation protocol, which employs top-1 accuracy alongside an answer-shuffling robustness criterion, and provide baseline results using state-of-the-art audio-language models (Qwen2-Audio-7B, AudioFlamingo 2, and Gemini-2-Flash). The non-archrival benchmark is released as an open-source resource and used for the DCASE 2025 AQA Challenge. Preliminary analyses on the development set reveal substantial variability in model performance across both tasks and domains. Ultimately, this benchmark is intended to foster progress in the auditory comprehension and reasoning abilities of audio-language models, capabilities that are critical for enabling systems to perceive, interpret, and interact with the acoustic world in a manner approaching human-level acuity.