Reproduce “ESResNet: Environmental Sound Classification Based on Visual Domain 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

Andrey Guzhov, Federico Raue, Jörn Hees, Andreas Dengel

This paper presents ESResNet models, which adapt computer vision's ResNet architecture and ImageNet pre-training for Environmental Sound Classification (ESC) using log-power spectrograms, establishing improved accuracy benchmarks (e.g., 91.5% on ESC-50, 85.4% on UrbanSound8K stereo) while also clarifying evaluation standards for the UrbanSound8K dataset by addressing reproducibility issues.

Key insights

  • Transfer learning from large visual datasets (ImageNet) provides substantial performance gains for ESC, particularly on smaller audio datasets, demonstrating the effectiveness of cross-domain adaptation.
  • Effectively utilizing multi-channel stereo audio via a Siamese-like architecture offers tangible performance improvements in environmental sound classification.
  • Strict adherence to official dataset splits is paramount for accurate and comparable benchmarking in ESC, as unofficial splits lead to artificially inflated performance metrics.

Summary by alphaXiv.

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