Audio recorder forensic identification in 21 audio recorders

Audio recorder forensic identification is to determine originating devices of questioned audio recordings. In this paper, statistical features both in time and frequency domains were studied and were used to represent and encode device-related information. Statistical techniques in fields of machine learning and pattern recognition were utilized to classify computed features into each originating audio recorder. More specifically, features of sampling histogram and spectral mean were extracted from near-silence segments of digital recordings which were recorded by using 21 audio recorders. A support vector machine classifier was used to learn device-related feature patterns upon one half of the complete set of audio samples, and then was evaluated by using the rest of data. The experimental results show a high accuracy of up to 96.72% in the correct classification of the 21 audio recorders. Our paper was the first study which was to study and evaluate the performances of different features for device-related information encoding in a relatively big audio recorder database. The proposed method would offer a practical guide for forensic identification of audio recorders.