COdec Recognition From Decoded Audio

Identifying or verifying the coding of an audio recording in uncompressed format may be useful in many fields, from audio authenticity assessment, in forensic science, to optimizing the use of biometric speaker recognition systems, when the test samples are collected remotely. In this paper, a channel recognition algorithm is presented for use with uncompressed audio, in the absence of noise. We considered the learning capabilities of Gaussian mixture models (GMM) and noticed the slightest variations induced by the transmission channel into the speech signal can be used to make channel recognition decisions. To the best of our knowledge, this is the first holistic approach of detecting adaptive multi-rate (AMR), wide band AMR (AMRWB) and G.729 codecs in an uncompressed audio, although these are some of the most used codecs in global systems for mobile telephony (GSM) and voice over Internet protocol (VoIP).