This paper suggests the nonlinear parameter of self-similarity as a novel feature to be employed in wavelet packet based voice signal analysis. Two groups of normal and pathological voice signals have been decomposed using wavelet packets. Next, self similar characteristics of reconstructed signals in each node have been calculated. Consequently, discrimination ability of each node has been obtained using Davies-Bouldin criterion. In the following, eight most discriminant nodes have been identified to construct feature vector parameters. To reduce the feature vector dimensionality Principal component analysis (PCA) has been employed. Finally, an artificial neural network has been trained to classify normal and pathological voices. The results show that self-similarity parameter can be a reliable feature in wavelet packet based voice signal analysis. Moreover, selected sub-bands are distributed over the whole available frequencies which shows that pathological factors do not influence specific frequency range which accentuates the role of WP decomposition.
Index Terms. vocal disorder, wavelet packet, selfsimilarity, Davies-Bouldin Criterion
Cite as: Khadivi Heris, H., Aghazadeh, B.S., Nikkhah-Bahrami, M. (2007) Classification of pathological voice signals using self-similarity based wavelet packet feature extraction and Davies-bouldin criterion. Proc. Models and Analysis of Vocal Emissions for Biomedical Applications (MAVEBA 2007), 85-88
@inproceedings{khadiviheris07_maveba,
author={H. {Khadivi Heris} and B. S. Aghazadeh and M. Nikkhah-Bahrami},
title={{Classification of pathological voice signals using self-similarity based wavelet packet feature extraction and Davies-bouldin criterion}},
year=2007,
booktitle={Proc. Models and Analysis of Vocal Emissions for Biomedical Applications (MAVEBA 2007)},
pages={85--88}
}