Reproduce “LEAN: Light and Efficient Audio Classification Network” 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
Shwetank Choudhary, CR Karthik, Punuru Sri Lakshmi, Sumit Kumar
Researchers at Samsung R&D Institute developed LEAN (Light and Efficient Audio Classification Network), a model designed for on-device audio classification that combines time-domain and frequency-domain feature extraction with cross-attention mechanisms. It achieves an mAP of 0.4677 on FSD50K and, after quantization, results in a 4.5MB model size, demonstrating a 22% improvement in on-device performance over a baseline Att-RNN model.
Key insights
- Temporal and frequency-domain features are complementary for audio classification, with cross-attention mechanisms effectively fusing them.
- Processing raw waveforms alongside spectral features captures dynamic patterns and transitions that might be missed by frequency-only approaches.
- Lightweight model architectures can achieve competitive performance suitable for on-device deployment by carefully balancing feature extraction and fusion methods.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.