1. Abstract

In this paper an edit detection method for forensic audio is proposed. It develops and improves a previous method introduced by the authors through changes in the signal processing chain and a novel detection criterion. As with the original method, Electrical Network Frequency (ENF) analysis is central to the novel edit detector, for it allows monitoring anomalous variations of the ENF related to edit events. Working in unsupervised manner, the edit detector compares the extent of ENF variations, centered at its nominal frequency, with a variable threshold that defines the upper limit for normal variations observed in unedited signals. ENF variations caused by edits in the signal are likely to exceed the threshold providing a mechanism for their detection.  The proposed method is evaluated in both qualitative and quantitative terms via two distinct annotated databases. Results are reported for originally noisy database signals as well as versions of them further degraded under controlled conditions. A comparative performance evaluation, in terms of equal error rate (EER) detection, reveals that, for the Carioca 1 database, an improvement from 7\% to 4\% EER is achieved, respectively from the original to the new edit detection method. When the signals in the databases are amplitude clipped or corrupted by broadband background noise, the performance figures of the novel method follow the same profile of those of the original method.

Download a post-print version of the paper:

DOI: http://dx.doi.org/10.1109/TIFS.2014.2363524

2. Demo 1: Voice Activity Detector (VAD)

• Demonstration with Matlab code

3. Demo 2: TPSW Filtering

• Demonstration with Matlab code
• Video Demonstration

4. Selection of Signals Featured in the Databases Used for Performance Evaluation (estimated SNR indicated in parenthesis)

• Audio clips from the Carioca 1 Database
  • Sample 1 (21.4 dB), Sample 2 (22.8 dB), Sample 3 (28.2 dB), Sample 4 (28.5 dB), Sample 5 (29.11 dB)
• Audio clips from the Spanish Database
  • Sample 6 (20 dB), Sample 7 (23.1 dB), Sample 8 (22.4 dB)

5.Entire Carioca 1 Database (300 MB)

• Download here.
• The Carioca 1 database has been produced by the Digital Signal Processing Laboratory - Department of Electrical Engineering -  Military Institute of Engineering (IME-RJ), Brazil.
• Authorization to use the Carioca 1 database must be obtained from Dr. José A. Apolinário Jr.
• When referring to the Carioca 1 database, please cite
  • D. P. N. Rodríguez, J. A. Apolinário Jr., and L. W. P. Biscainho, “Audio authenticity: Detecting ENF discontinuity with high precision phase analysis,” IEEE Transactions on Information Forensics and Security, vol. 5, no. 3, pp. 534–543, Sept. 2010.

6. Matlab Codes for Generating the Figures shown in the Paper

• Package with matlab codes

7. Edit Detection in Signals with Weak ENF Component

The strength of the ENFC (at 60 Hz) is weak in the 3 test signals below, especially in test_sig3.wav, as the log-spectrograms (from Audacity) reveal.

The edit detection results (same setup as described in the paper) for the 3 signals above are given below:

The vertical dashed lines indicate the cut or insertion points. For test_sig1.wav, both edit points of the insert are detected correctly. For test_sig2.wav, the second edit point of the insert is detected. The other peaks above the threshold are not considered, since they are inside voice-active parts of the signal. For test_sig3.wav, the cut point is correctly detected, despite the weak ENF.

The test signals and the simulation codes are available for verification from the link below

http://www.lncc.br/~pesquef/TIFS2014/demo_weak_ENF.zip