Reproduce “MMAC: A Massive Multi-dimensional Benchmark for Audio Captioning” 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

Weijie Wu, Junbo Li, Lin Li, Jun Fang, Qingyang Hong

Researchers from Xiamen University and DiDi Global Inc. introduce MMAC, a massive multi-dimensional benchmark for audio captioning, offering a diagnostic framework to evaluate Audio Large Language Models by assessing information coverage and description reliability across 6 capability categories and 15 fine-grained dimensions.

Key insights

  • Advanced AudioLLMs exhibit varying trade-offs between information coverage and description reliability across different audio understanding dimensions, with high coverage not necessarily correlating with high precision.
  • Supervised fine-tuning for captioning can significantly improve information coverage but does not consistently enhance precision, particularly for complex dimensions like inferring implicit meanings.
  • Increasing the output length of generated captions generally leads to greater information coverage but often results in a decrease in description reliability (precision).

Summary by alphaXiv.

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