Sound Event Detection and Time-Frequency Segmentation from Weakly Labelled Data
share
Sound event detection (SED) aims to detect what and when sound events happen in an audio clip. Sound events can be segmented in the time-frequency (T-F) domain and is called T-F segmentation. Many supervised SED algorithms rely on strongly labelled data which contains labels of onset and offset times of sound events. However, many audio tagging datasets are weakly labelled, that is, only the presence or absence of the sound events is known, without knowing their onset and offset times. In this paper, we propose a SED and T-F segmentation framework trained with weakly labelled data. In the training stage, we propose a segmentation mapping applied on a T-F representation of an audio clip to obtain T-F segmentation masks of sound events. We then apply a classification mapping on each T-F segmentation mask to estimate the presence probability of each sound event. Both of the segmentation mapping and classification mapping are trained jointly. In T-F segmentation, T-F segmentation masks can be obtained by presenting a T-F representation of an audio clip to the trained segmentation mapping. In SED, predicted onset and offset times can be obtained from the T-F segmentation masks. We propose to model the segmentation mapping using a convolutional neural network and the segmentation mapping using a global weighted rank pooling (GWRP). As a byproduct, separated waveforms of sound events can be obtained from their corresponding T-F segmentation masks. Experiments on the remixed DCASE 2013 dataset show that the proposed method obtains an area under the curve (AUC) score of 0.948 in audio tagging and 0.893 in sound event detection, outperforming a deep neural network baseline of 0.719 and 0.616, respectively.
READ FULL TEXT