Accuracy of Divergence Measures for Detection of Abrupt Changes

Numerous divergence measures (spectral distance, cepstral distance, difference of the cepstral coefficients, Kullback-Leibler divergence, distance given by the General Likelihood Ratio, distance defined by the Recursive Bayesian Changepoint Detector and the Mahalanobis measure) are compared in this study. The measures are used for detection of abrupt spectral changes in synthetic AR signals via the sliding window algorithm. Two experiments are performed; the first is focused on detection of single boundary while the second concentrates on detection of a couple of boundaries. Accuracy of detection is judged for each method; the measures are compared according to results of both experiments.

Authors:

• P. Bergl

Keywords:

• Abrupt changes detection
• autoregressive model
• divergence measure.

References:

[1] U. Appel, "A comparative study of three sequential time series segmentation
algorithms," Signal Processing, vol. 6,pp. 45 − 60, 1984.
[2] U. Appel, "Adaptive Segmentation of Piecewise Stationary Time Series,"
Information Sciences, no.29,pp. 27 − 56, 1983.
[3] J.O. Ruanaidh, W.J. Fitzgerald, Numerical Bayesian Methods Applied
to Signal Processing, Statistics and Computing, Springer, Berlin, 1996.
[4] R. Cmejla, P. Sovka, "Recursive Bayesian Autoregressive Changepoint
Detector for Sequential Signal Segmentation," in EUSIPCO-2004 -
Proceedings [CD-ROM]. Wien: Technische Universitat, 2004.
[5] J.J. Sooful, J.C. Botha, "An acoustic distance measure for automatic
cross-language phoneme mapping," PRASA-01, pp. 99 − 102, South
Africa, November 2001.
[6] L. Couvreur, J.M. Boite, "Speaker Tracking in Broadcast Audio Material
in the Framework of the THISL Project," Proc. of ESCA ETRW
Workshop on Accessing Information in Spoken Audio, Cambridge (UK),
pp. 84 − 89, April 1999.