Highly accurate non-intrusive speech forensics for codec identifications from observed decoded signals

The ability to detect a particular speech codec from only the decoded audio has several useful forensic and system performance improvement applications. This paper presents a novel scheme for non-intrusive identification of speech codecs. The identification approach is based upon comparing a profile of a set of noise spectra and a time-domain histogram from the decoded speech to those from the candidate codecs. The presented results show a very high accuracy in identifying speech contemporary codecs from a diverse set of types and encoding rates. The presented codec identification scheme has a very low misidentification rate, including in the high coding rate regime where it improves on previous works by achieving perfect identification. This performance is achieved while reducing the duration of the analysis window of speech from 2 minutes to only 4 seconds.