Detection of AMR double compression using compressed-domain speech features

Abstract

The adaptive multi-rate (AMR) audio codec is an established standard of speech signal compression, for either transmitting speech signal over mobile networks or storing digital audio on handheld devices. The widespread use of AMR codec and the high availability of tampering software have increased authentication cases in court. AMR double compression detection is a challenging engineering problem and a topic of multimedia forensics. As a general rule, a double compressed AMR file cannot be considered an original file. In this paper, a new method based on support vector machine (SVM) is proposed to classify single and double compressed AMR digital audio. Instead of using the decoded speech waveform, the proposed method uses only compressed-domain speech features. Specific parameters are extracted from encoded AMR files and used to create a set of statistical features. After applying robust scaling to features and selecting the SVM model, recursive feature elimination with correlation bias reduction (RFE-CBR) is used to determine the best number of features to maximize accuracy in SVM classification. The experiments reveal that the proposed algorithm can discriminate single and double compressed AMR speech, outperforming the published methods. The average accuracies using TIMIT database and CARIOCA1 database, the last recorded from landline phone calls, are about 99%. Other experiments, including frame offset attack and noise addition, found that the method is robust and reliable.