Semi-Supervised NMF-CNN for Sound Event Detection

The lack of strongly labeled data can limit the potential a Sound Event Detection (SED) system trained using deep learning approaches. To address this issue, paper proposes novel method to approximate strong labels for weakly Nonnegative Matrix Factorization (NMF) in supervised manner. Using combinative transfer and semi-supervised framework, two different Convolutional Neural Networks (CNN) are synthetic data, approximated unlabeled where one model will produce audio tags. In contrast, other frame-level prediction. proposed methodology is then evaluated on three subsets Classification Acoustic Scenes Events (DCASE) 2020 dataset: validation dataset, challenge evaluation public YouTube dataset. Based results, our outperforms baseline by minimum 7% across these subsets. addition, also top 3 submissions from DCASE 2019 task 4 datasets. Our performance competitive against submission data. A post-challenge analysis was performed which revealed causes difference between 4. leading that we observed 1) detection threshold tuning 2) augmentation techniques used. We could perform better than first place 1.5% changing method. more robust long-duration clips, outperformed them 37%.