Details
Last Updated on Wednesday, 21 March 2018 14:42
Published on Monday, 25 May 2015 13:48
Written by Paulo Esquef
Hits: 2676


Companion Web-page of the Paper "Improved Edit Detection in Speech via ENF Patterns", published in the proceedings of the 7th IEEE International Workshop on Information Forensics and Security (WIFS) 2015.

Authors: Paulo A. A. Esquef1, José A. Apolinário Jr.2, and Luiz W. P. Biscainho3

1Coordination of Systems and Control,
-------
Military Institute of Engineering, Rio de Janeiro, Brazil
--------
3Signals, Multimedia and Telecommunications Laboratory
Federal University of Rio de Janeiro & PEE/COPPE UFRJ, Rio de Janeiro, Brazil
--------

1. Abstract

Abstract—In a recent paper published in the IEEE TIFS, we proposed an edit detection method based on the instantaneous variations of the Electrical Network Frequency (ENF). In this work we modify the detection criteria of that method by taking advantage of the typical pattern of ENF variations elicited by audio edits. We describe the implemented modifications and directly confront the performance of both methods using two distinct signal databases that contain real-life speech recordings. The experimental results demonstrate that the new proposition has an improved performance in terms of lower equal error rates when compared to its former version.

2. Related demos from the IEEE TIFS paper DOI: http://dx.doi.org/10.1109/TIFS.2014.2363524

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

4. Matlab Code for Replicating the Results with Amplitude Clipping for the Carioca 1 Database.

5. Suplementary Results

  • Detector Error Tradeoff Curves of the proposed method for the Carioca 1 Database and its version degraded by amplitude clipping at 0.2%.

Setup 1: Original Carioca 1 database signals, λ=0.7 (fixed), FP and FN measured for several values of G (indicated near the diamond marks).


 

 

 

 

 

 

 

 

 

 

Setup 2: original Carioca 1 Database signals, G=3.3 (fixed), FP and FN measured for several values of λ (indicated near the diamond marks).

Setup 3: Carioca 1 Database signals with amplitude clipping at 0.2%, λ=0.7 (fixed), FP and FN measured for several values of G (indicated near the diamond marks).

Setup 4: Carioca 1 Database signals with amplitude clipping at 0.2%, G=16.9 (fixed), FP and FN measured for several values of λ (indicated near the diamond marks).

The above DET curves confirm the expected behavior: since G controls the overall height of d[n], increasing G (for a fixed λ) tends to increase the rate FN and decrease that of FP. Since setting λ near 1 implies requiring the observed pattern in d[n] around a detected local maxima to be very similar to the templates, increasing λ toward 1 will tend to increase FN and decrease FP.

  • Performance of the proposed detection method as a function of λ for the Carioca 1 Database corrupted with LP noise.

Carioca 1 database, LP Noise, SNR 5dB

λ

 G_opt mean EER (%)
0.725 20.6 46.9
0.750 18.3 45.9
0.775 15.6 47.8
0.800 12.6 47.3
0.825 8.6 47.9
0.850 5.6 51.2
0.875 1.9 50.1
Carioca 1 database, LP Noise, SNR 10dB
λ G_opt mean EER (%)
0.725 19.6 48.6
0.750 17.5 48.6
0.775 14.5 49.5
0.800 11.5 50.0
0.825 8.4 49.0
0.850 5.2 50.6
0.875 2.0 49.5
Carioca 1 database, LP Noise, SNR 15dB
λ G_opt mean EER (%)
0.725 13.4 43.0
0.750 11.6 42.3
0.775 8.5 45.2
0.800 6.5 44.5
0.825 4.6 44.6
0.850 2.3 45.5
0.875 0.8 44.5
Carioca 1 database, LP Noise, SNR 20dB
λ G_opt mean EER (%)
0.725 10.5 34.5
0.750 7.8 35.0
0.775 6.6
 35.5
0.800 4.7 35.4
0.825 2.3 36.9
0.850 0.9 38.5
0.875 - -