Study Of Attention-based Deep Models For Acoustic Scene Classification

Helping devices understand the environment where the devices are located via the analysis of sound is the main target in the field of machine auditory research.Machine auditory is a field of research that involves computational auditory scene analysis.Machine auditory system needs to perform similar processing tasks as the human auditory system and is part of a broader research topic in related fields such as machine learning,robotics and artificial intelligence.The acoustic scene classification problem is a sub-problem of computational auditory scene analysis and one of the most difficult tasks.People can understand where they are,such as busy streets,offices,etc.,and they can identify independent sound events as well,such as passing cars and footsteps.One of the purposes of acoustic scene classification task is to simulate the ability of humans to distinguish the acoustic scene:classification of the acoustic scene described by a given audio clip.Nowadays,wearable smart devices and smart home are playing an important role in people's daily life.Understanding the environment by acoustic analysis can help these devices smarter.Therefore,the acoustic scene classification has been receiving increasing attention.In this paper,we designed a series of deep learning models including convolutional neural network,recurrent neural network,bilinear model and capsule network model.The convolution neural network model based on attention mechanism achieved the classification accuracy of 87.9%,which is better than the best result of 86.9%in the single model compared with all submitted results of DCASE2016.To the best of our knowledge,this is the first try of attention mechanism in acoustic scene classification.In order to further exploiting the advantages of deep learning,this paper designed a deep convolutional neural network model based on transfer learning,using the pre-trained model on large dataset of other filed to make up the lack of training data in the field of acoustic scene clas-sification.This model achieved the accuracy of 90.3%,higher than the first place of 89.7%in DCASE2016.