Identification of audio codecs from single coded and transcoded audios based on coding properties

The research presented in this paper provides a novel technique to identify the codec of an encoded audio. Since the technique does not perform decoding, utilize any coding metadata, or assume information about the structure describing the bit stream format of a codec, it is simple and fast. The main idea of the technique is that design principles of a codec such as compression level, audio quality and complexity will reveal themselves on the coded audio. To exploit this, the technique samples 2-4 kilobytes of data from a coded audio and analyzes the randomness and chaotic nature of the sampled data to build statistical models that represent encoding process associated with different codecs. To evaluate the effectiveness of the technique in identification of the codec, in experiments we use 16 of the most popular audio codecs used for high quality audio compression and in PSTNs, cellular networks, and VoIP networks by setting encoding parameters of each codec to its most commonly used values. Tests are performed on both speech and music data sets to eliminate any effect which may stem from contents of audios. Results show that the codec of a singly encoded audio can be identified with an accuracy of more than %95. Several transcoding scenarios were also performed. Those results show that the scheme can even identify the first encoder of a doubly-encoded audio with an accuracy range of around %80 to %90 or more as long as the second codec operates on higher bit rates than the first one.