Reproduce “SELD-Mamba: Selective State-Space Model for Sound Event Localization and Detection with Source Distance Estimation” 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
Da Mu, Zhicheng Zhang, Haobo Yue, Zehao Wang, Jin Tang, Jianqin Yin
Researchers at Beijing University of Posts and Telecommunications introduced SELD-Mamba, an architecture leveraging selective state-space models for Sound Event Localization and Detection (SELD) with Source Distance Estimation (SDE). The model achieves a SELD score of 0.381 with 75.14M parameters and 6.35 G/s MACs, demonstrating competitive performance with significantly higher efficiency than Transformer-based counterparts.
Key insights
- Selective state-space models like Mamba can replace Transformer self-attention in SELD tasks, offering comparable or superior performance with substantially reduced computational overhead.
- A two-stage training strategy effectively optimizes multi-task SELD performance by addressing tasks of varying difficulty, notably improving SDE accuracy after core localization and detection capabilities are established.
- The combination of log-mel spectrograms as input features for the SDE branch, without intensity vectors, yields improved overall SELD performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.