Acoustic Scene Online Detection System Based On LSTM Network

With the development of the Internet,the promotion of multimedia technology and the spread of mobile devices,there have been a large number of online platforms with live broadcast as their main business.The applications that keep sharing videos as their main business are becoming more popular and their target users are becoming more extensive.The large amount of multimedia data and the variety of them,carried by the platforms and applications,have brought great inconvenience to the users to search for the desired data,and also cause trouble for the supervision of the network.Aiming at the online audio data carried by multimedia applications mentioned above,we set up an online detection system that can judge the class of an audio within the acoustic scene classes.This paper will introduce three main concerns:Firstly,an acoustic scene recognition method based on Gaussian Mixed Model(GMM)is studied.We briefly introduce the DCASE2016 Task1 acoustic scene recognition competition and the acoustic scene recognition baseline system based on GMM.The baseline system uses the Mel Frequency Cepstrum Coefficient(MFCC)as features.A GMM model is trained for each acoustic scene class.In the classification step,the maximum value is selected from the outputs of GMM models,and the class,corresponds with the maximum value,is considered as the predicted category of the data.Secondly,an acoustic scene recognition method based on Long Short Term Memory Network(LSTM)is proposed.This method distinguishes the audio streams with acoustic scene classes,using the acoustic scene recognition model based on LSTM network.On the basis of introducing the characteristics of LSTM network structure,we discuss the process,related to the model,of audio preprocessing,feature extraction,segmentation processing,training and testing.And then,the test results of the recognition model are given.Using the competition data set of the DCASE2016 Task1 acoustic scene recognition,the method achieves a maximum accuracy of 81.8%.Finally,an online detection system is designed and implemented.The category of acoustic scene in online audio data is an open set,and the offline training data cannot contain all categories in online data.Therefore,this paper proposes a system framework that can dynamically distinguish whether there is new category data in the online data,and can analyze the new category after the retraining.For this purpose,a GMM discrimination model and an acoustic scene recognition model based on LSTM are designed.At the end,the test results of the system show that the recognition model based on LSTM can classify the acoustic scene online,and the new acoustic scene category can be accepted and recognized by the system.