Acoustic Scene Classification Based On Adversarial Domain Adaptation

Acoustic Scene Classification(ASC)aims to capture scenes described by processing and analyzing acoustic signals.The development of this field makes the data set continuously supplemented and enriched.A variety of data sets means the inconsistency of data distribution,which will further affect the performance of acoustic scene classification.This paper proposes an acoustic scene classification method based on adversarial domain adaptation,which reduces the impact of inconsistent data distribution on the performance of ASC through feature alignment.Firstly,the cross-device acquisition causes some distortions of audio signals,resulting in inconsistent data distribution.To solve this problem,this paper proposes a conditional adversarial domain adaptation method based on entropy weighting and guidance information.In this method,the data distribution captured by different devices is aligned by adversarial training to extract device-independent features and reduce the interference of device discrepancy to the classification model.Compared with the traditional adversarial domain adaptation,our method takes the scene information as the condition of feature alignment so that the feature distribution is aligned according to scenes.Entropy weighting matches different weights with samples to improve feature alignment.The guidance information makes the scene information and entropy weighting generation more appropriate.On the dataset of ASC with multiple devices in DCASE2019,the proposed method is more effective than traditional adversarial domain adaptation methods to avoid inconsistencies in distribution caused by device diversity,with an average accuracy of 61.9%.Besides,data distributions of undefined scene audio and defined scene audio are inconsistent due to the difference in semantic information or sound events.To solve this problem,an open set ASC method based on adversarial domain adaptation is proposed in this paper.The proposed method uses a preset pseudo-threshold to program the boundary between known scenes and unknown scenes and filters out unknown scenes through adversarial training of the two kinds of scenes to complete the classification task of known scenes.The proposed method proves that the adversarial domain adaptation can filter out the undefined or uninterested scenes and classify known scenes on the dataset of open set ASC in DCASE2019.Compared with the baseline,the accuracy is improved by 12.4%in known scenes and has an improvement of 17.5%in unknown scenes.Finally,considering that the difference of devices and the diversity of recording scenes may occur at the same time and lead to the inconsistency of data distribution,this paper proposes a new application scenario:open set multi-device ASC and a solution based on adversarial domain adaptation.The proposed method uses feature alignment to extract device-independent feature representations to classify known scenes from multiple devices and recognize unknown scenes through adversarial training.The effectiveness of the model is verified on the joint dataset of DCASE2019 multi-device ASC and open-set ASC.To solve the problem of inconsistent data distribution caused by device diversity or unknown scenes,this paper proposes adversarial domain adaptation for multi-device,open set,and open set multi-device ASC.This method reduces the impact of inconsistent distribution on scene classification.