Audio authentication using the kurtosis of ESPRIT based ENF estimates

Digital audio recordings are an important source of evidences related to various crimes and conflicts. The authentication of this source of evidence is often a necessary and critical task, but still subject to many challenges. In order to identify audio tampering we present a new technique to detect deletions and insertions of audio snippets by exploiting anomalous variations in Electric Network Frequency (ENF) estimates of interfering power grid signal in an audio recording under test. The method is based on the hypothesis that insertions and deletions of audio snippets produce phase discontinuities in ENF. Such discontinuities cause abnormal disturbances on the estimated ENF. First we employ an ESPRIT based ENF estimation technique. Next we propose a feature based on the kurtosis of the ESPRIT estimates that measures the outlierness of ENF variations. Finally we propose an automatic detector of ENF disturbance by a linear discriminant. The proposed method outperforms state-of-the-art approach in low signal-to-nose rates scenarios. To asses our results we use a corpus with 100 edited and 100 unedited authorized audio recordings of phone calls named Carioca 1.