Reproduce “Improving Sound Event Detection In Domestic Environments Using Sound Separation” on OpenResearch

About this paper

Nicolas Turpault, Scott Wisdom, Hakan Erdogan, John Hershey, Romain Serizel, Eduardo Fonseca, Prem Seetharaman, Justin Salamon

Integrating universal sound separation as a pre-processing step improves Sound Event Detection (SED) performance in domestic environments. A late integration strategy, combining predictions from both the original mixture and separated sources, achieved a 40.7% F1-score and 0.570 PSDS on the REC EVAL set, a modest improvement over the DCASE 2020 Task 4 baseline.

Key insights

• Late integration, which applies SED independently to the mixture and separated sources and then combines their predictions via power mean aggregation, is consistently the most effective strategy.
• Generically trained sound separation models (on dry FUSS data) yielded better SED performance on real-world recordings compared to reverberant FUSS-trained models or models trained on more specific synthetic DESED+FUSS data, highlighting challenges in domain generalization.
• Intermediate values for the power mean aggregation parameters (`p` and `q`) in late integration are optimal, indicating that a balanced contribution from both the original mixture and separated components is crucial for best performance.

Summary by alphaXiv.