Deep and shallow feature fusion and recognition of recording devices based on attention mechanism

Abstract

In the device source identification task, there is a lot of redundant information in the traditional MFCC, GSV and i-vector features, which affects the accuracy of device source identification. In this paper, in order to extract the key information from the device source, we propose a multi-feature fusion attention mechanism network to achieve improved accuracy. First, the multi-feature fusion attention mechanism network we proposed directly compresses the MFCC, GSV and i-vector features that characterize the source of the device using convolutional neural networks, reducing the dimension and computation of the feature vector and improving the feature vector Representativeness Second, we use a feature attention mechanism to adaptively rescale feature characteristics by considering the interdependence between features. Finally, we also use a combination of deep learning and machine learning. The back-end uses SVM models for classification. Extensive experiments show that our proposed method achieves better accuracy in device source recognition tasks.