DOUBLE COMPRESSED AMR AUDIO DETECTION USING LINEAR PREDICTION COEFFICIENTS AND SUPPORT VECTOR MACHINE
The adaptive multirate codec (AMR) standard usage has been intensive in mobile networks as well as a speech stor-age format. Due to its high availability, many AMR audio files take on a forensic evidence condition, which implies the need to demonstrate they are authentic. Detecting AMR double compression means that, in the multimedia forensics context, the file is not an original one and a tampering likely happened, since it is always necessary to decode and encode again to change utterance meaning. In this paper, we show a new method to detect AMR double compression based on compressed-domain linear predic-tion (LP) coefficient extraction, statistical feature computation from LP coefficients and SVM application. The experiments demonstrate the proposed method can discriminate double compressed AMR files from single compressed files with satisfactory accuracy, either using mixed first compression bitrate sets, or fixed first compression bitrate ones. Using the TIMIT corpus, the average accuracy reached 93.66 %, which is a very satisfactory result.