Combining Multi-Scale Features Using Sample-level Deep Convolutional Neural Networks for Weakly Supervised Sound Event Detection

Abstract

This paper describes our method submitted to large-scale weakly supervised sound event detection for smart cars in the DCASE Challenge 2017. It is based on two deep neural network methods suggested for music auto-tagging. One is training sample-level Deep Convolutional Neural Networks (DCNN) using raw waveforms as a feature extractor. The other is aggregating features on multi-scaled models of the DCNNs and making final predictions from them. With this approach, we achieved the best results, 47.3% in F-score on subtask A (audio tagging) and 0.75 in error rate on subtask B (sound event detection) in the evaluation. These results show that the waveform-based models can be comparable to spectrogram-based models when compared to other DCASE Task 4 submissions. Finally, we visualize hierarchically learned filters from the challenge dataset in each layer of the waveform-based model to explain how they discriminate the events.