GMM-based audio codec detection with application in forensics
Dragoş Drăghicescu, Gheorghe Pop, Dragoş Burileanu, Corneliu Burileanu · 2015
Reliable detection of previous compression in audio is a sensible subject of interest for both forensic analysts and audiophiles. The former use could lead for audio authentication, while the latter may be able to detect simulated high quality of audio. Two state-of-the-art machine learning solutions, Gaussian Mixture Models (GMMs) with Universal Background Model and Gaussian Super-Vectors with Support Vector Machines (GSV-SVM), were used to learn distinctive feature statistics and decide upon which codec was used. The method we propose seeks to detect traces of most widely used audio codecs, from landline or mobile telephone recordings, to more general compression currently used in most portable recorders. A frequency range of 0 to 4 kHz was targeted in our investigation, although our approach can be readily applied at higher frequencies.