Research On Acoustic Scene Classification Based On Convolutional Neural Network

The task of acoustic scene classification is to classify the input audio into a predefined acoustic scene class.Its purpose is to simulate the ability of human beings to distinguish the acoustic scene through the machine.At present,the acoustic scene classification technology has been applied in many fields,such as ecological environment monitoring,public security intelligent monitoring,speech recognition technology,automatic driving assistance and so on,which also makes the acoustic scene classification task become an important field of the development of artificial intelligence.In recent years,deep learning technology has a good performance in many fields of artificial intelligence,and convolutional neural network has gradually become the most popular method of acoustic scene classification task.In this thesis,we explore the impact of different convolutional neural network architectures on the classification results in the acoustic scene classification problem,and on this basis,we propose the multi-task learning method(primary task and auxiliary task)of acoustic scene classification combined with sound event detection.The main work includes the following two aspects:(1)The construction method of convolution neural network model for acoustic scene classification is studied.Due to the particularity of acoustic scene classification task,deep structure convolution neural network can not achieve good results.In this thesis,the acoustic scene classification problem is studied by adjusting the number of convolution layers,transition layers and network architecture types from general architecture of convolution neural network.Finally,the model constructed by us has been widely tested on TUT acoustic scene 2017 dataset,compared with the DCASE2017 baseline system,our model has significantly improved the classification accuracy.(2)A method of acoustic scene classification based on multi-task convolution neural network is proposed.Due to the particularity of acoustic scene classification task and the limitation of dataset,it is difficult to improve the accuracy by designing the model architecture.For this reason,this thesis proposes a multi-task learning method,which combines the acoustic scene classification task with the sound event detection task by using their correlations.The experimental results show that the classification accuracy of the convolutional neural network model based on multi-task learning is better than that of the convolutional neural network model based on single task.The segmentation error rate of sound event detection is higher than that of baseline system,but the accuracy of F-score is improved.The experimental results verify the effectiveness of our proposed method.