Reproduce “One Waveform, Two Stories: Benchmarking Speech-Acoustic Conflict in Audio-Language Models” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Wenbin Wang, Yuheng Li, Koichiro Yoshino

SAC-Bench tests whether audio-language models use environmental sound when speech makes a competing claim. It pairs environmental recordings with compatible or conflicting synthesized speech across seven tasks covering event presence, identity, timing, and spatial properties. Its 616,031 clips span 188 sound-event classes. Qwen3-Omni Event Identity accuracy falls from 83.0% with compatible speech to 39.8% with conflict; these are condition-level results.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.