Research On Weakly Labeled Polyphonic Sound Event Detection Based On Deep Learning

As one of the most common information carriers in human life,sound has the characteristics of convenient collection,small storage space,and no limitation of light.Therefore,how to extract usable information from sound and apply it in real life has become a research hotspot that has received much attention.Polyphonic sound event detection refers to classifying and locating the sound events in the audio recording,not only to give the category of each sound event,but also to give the onset and offset time of the event.Polyphonic sound event detection technology has a wide range of real-life applications,but the field started late,and the current detection technology is not yet fully mature.In the field of polyphonic sound event detection,there are two main problems that events overlap each other in the temporal domain and the lack of strongly annotated datasets.Aiming at the above two problems that need to be solved urgently,this paper studies a weakly labeled polyphonic sound event detection method based on deep learning.Capsule network is a newly proposed neural network architecture in recent years,which has achieved rapid development in image recognition,target detection and other directions.In this paper,a polyphonic sound event detection method based on capsule network is proposed.Due to the large amount of parameters of the capsule network and the slow running speed,this paper introduces a non-iterative self-attention routing,which uses the multi-path primary capsule structure to perform detection on the weakly labeled dataset.The experimental results show that the method can be fast,accurate and effective detection.In addition,this paper also organically combines capsule network and recurrent neural network to realize polyphonic sound event detection.The dynamic routing algorithm with weight sharing is used,which can not only reduce the training parameters of the capsule layer significantly,but also reduce the training time of the model,at the same time,advanced detection performance is achieved by combining parallel early fusion convolution layer and recurrent neural network.