In pattern recognition or classification

extracting effective classification features from original pattern signals is very important.But

for a great number of non-stationary or time-varying signals

such as speech

radar

earthquake signals

etc.

classification features are often localized both in time and frequency

so thus extracting effective features from them by general transformation methods is very difficult.Wavelet packet transform can provides an arbitrary time-frequency decomposition for the signals

because a wavelet packet library contains many wavelet packet bases

which can handle the different components of a signal.Therefore

by selecting a suitable basis

the effective features can be extracted.This paper is mainly concerned with extracting effective features from the recognized or classified signals by selecting wavelet packet basis via given training sample sets.Three separability criteria

i.e.

distance criterion

divergence criterion and entropy criterion

are used for selecting the best basis.The performance of features extraction by wavelet packet transform is compared with that by wavelet transform through experiments.