| With the rapid development of intelligent speech technology and the rise of artificial intelligence related applications,acoustic scene classification(ASC)has been gradually applied to people’s daily life.It uses audio signal processing and deep learning technology to complete the recognition and classification of acoustic scenes(home,park,street scenes...),So as to achieve the purpose of identifying the surrounding environment.Aiming at the acoustic scene classification problem,this paper studies from three aspects,which are acoustic scene classification based on traditional acoustic representation,end-to-end acoustic modeling and fusion strategy based on multi feature system.In the aspect of traditional acoustic representation,this paper builds a sound field classification system based on ivector,and uses the generative representation ivector to further train the support vector machine to improve the distinctiveness of scene features;in the aspect of end-to-end acoustic modeling,this paper designs several groups of ASC systems,respectively using x-vector technology based on time-delay neural network,convolutional neural network and residual network,and compared with the traditional model,the performance of the acoustic scene classification system is further improved;in the aspect of system fusion strategy,this paper proposes a linear logistic regression algorithm to fuse score for multiple robust subsystems.Because different model structures will capture acoustic features from different angles,there are discriminative and complementary information between models.Aiming at the task of DCASE2019,this paper proposes a strategy of combining traditional generative representation with discriminative deep representation to learn the complementary information among multiple representations.Furthermore,aiming at the problem of sparse training data and mismatching of recording equipment in DCASE2020 international evaluation task,this paper proposes Data Augemnt algorithm like Mixup,Spec Augment and Device Augment to improve the performance of ASC system.The algorithm studied in this paper is verified on the corresponding tasks of DCASE,an international evaluation task of acoustic scene classification and detection.In the evaluation task of DCASE2019,the classification accuracy of the acoustic scene classification system based on traditional acoustic characterization and the end-to-end acoustic modeling system are 63.97% and 67.53% respectively,which are 1.47% and5.03% higher than the official baseline system.On the evaluation task of DCASE2020,the performance of the proposed algorithm achieves 70.24% classification accuracy,which is 16.44% higher than the baseline.In addition,the multi feature fusion strategy proposed in this paper is verified by a large number of experiments on DCASE evaluation tasks in 2019 and 2020 respectively.The results show that the traditional production acoustic representation and deep differentiated representation are highly complementary,and the fusion in both feature domain and score domain significantly improves the performance of the final acoustic scene classification system.In addition,the proposed fusion system has achieved excellent results in the evaluation of DCASE2020-Task1 A. |
