Audio Recorder Identification Using Reduced Noise Features

摘要

In this paper, we propose an audio recorder identification method as one of digital forensic technologies, where Wiener filter is used to extract noise sounds of recorders and their features are extracted by MIRtoolbox. A recorder identification model is generated by training SVM with the extracted noise features. To improve the identification performance, a new feature reduction method which uses inter-classes standard deviations of features is adopted. The experimental results for 11 audio recorders show 1% improvement over the method with no feature reduction. The improvement is not too noticeable as expected, but the number of features can be reduced up to one third of the method with no feature reduction. The results also show that the proposed feature reduction method is competitive over the other well-known methods such as PCA, LDA and R-squared.