Large-scale acoustic scene analysis with deep residual networks

Abstract

Many of the recent advances in audio event detection, particularly on the AudioSet dataset, have focused on improving performance using the released embeddings produced by a pre-trained model. In this work, we instead study the task of training a multi-label event classifier directly from the audio recordings of AudioSet. Using the audio recordings, not only are we able to reproduce results from prior work, we have also confirmed improvements of other proposed additions, such as an attention module. Moreover, by training the embedding network jointly with the additions, we achieve a mean Average Precision (mAP) of 0.392 and an area under ROC curve (AUC) of 0.971, surpassing the state-of-the-art without transfer learning from a large dataset. We also analyze the output activations of the network and find that the models are able to localize audio events when a finer time resolution is needed. In addition, we use this model in exploring multimodal learning, transfer learning, and realtime sound event detection tasks.