Discriminant wavelet basis construction for speech recognition

In this paper, a new feature extraction methodology based on Wavelet Transforms is examined, which unlike some conventional parameterisation techniques, is flexible enough to cope with the broadly differing characteristics of typical speech signals. A training phase is involved during which the final classifier is invoked to associate a cost function (a proxy for misclassification) with a given resolution. The sub spaces are then searched and pruned to provide a Wavelet Basis best suited to the classification problem. Comparative results are given illustrating some improvement over the Short-Time Fourier Transform using two differing subclasses of speech.