Weakly Supervised Sound Event Recognition On Noisy Label Dataset

Machine hearing refers to the ability of a machine to obtain information by analyzing sound signals.With the coming of the intelligent era,machine hearing has been paid more and more attention.One of the most important tasks is sound event recognition(SER).Sound event recognition refers to the interpretation of event information by analyzing audio signal.At present,the mainstream SER methods are based on deep learning,and there are two problems: first,deep learning relies on labeled data,and the labeling cost of sound events is very high,so it is difficult to obtain enough labeled data,which limits the performance of supervised learning model;second,the data of rare sound event types are difficult to obtain.In the above situation,the dataset can not provide enough supervision in the training process,so weak supervised learning method is needed to deal with this situation.To solve the above problems,this paper selects two sub-tasks of sound event detection task and machine abnormal sound detection task,and studies the corresponding weak supervised learning method.In the task of sound event detection,this paper first optimizes the structure of the model,and introduces residual connection based on convolutional recurrent neural network.Then,the importance of multi-scale feature fusion is analyzed,and the multi-resolution feature generation layer is proposed to generate multi-resolution features,and the dense connected recurrent neural network structure is proposed to fuse features with different levels of abstraction.At the same time,mean teacher model is applied to using unlabeled and weakly labeled data in the dataset to make the model learn better feature representation.Finally,combined with the above methods,the performance of the proposed system is improved by 13.4% compared with the baseline system.In the task of machine abnormal sound detection,this paper first selects two classification tasks: machine type recognition and machine number recognition,designs an abnormal sound detection model based on classification confidence,and uses mixup algorithm to smooth the feature space.Then,this paper proposes the unknown sample classification task,which is added to the model training as an additional task.This paper designs four methods to generate unknown samples.Then,this paper simulates the frequency response difference between devices in the actual environment,and verifies the effectiveness of the proposed method.Finally,compared with the baseline system,the proposed method achieves a performance improvement of 12.8%,and achieves a performance improvement of 25.8% with the frequency response difference between devices.