Parallel implementation for wavelet dictionary optimization applied to pattern recognition ? Leandro D. Vignolo 1 , Diego H. Milone 1,2 , Hugo L. Rufiner 1,2 y Enrique M. Albornoz 1 1 Grupo de Investigaci´ on en Se˜ nales e Inteligencia Computacional. Facultad de Ing. y Cs. H´ ıdricas, Universidad Nacional del Litoral, Argentina 2 Laboratorio de Cibern´ etica, Facultad de Ingenier´ ıa, Universidad Nacional de Entre R´ ıos, Argentina {ldvignolo, dmilone, lrufiner, emalbornoz}@fich.unl.edu.ar Abstract. By means of full wavelet packet decomposition a redundant set of coefficients is obtained. For signal classification it is convenient to find a subset of these coefficients minimizing the error rate of a classifier. A problem arises because of the computational cost of GA solution. This work presents the parallelization of a genetic algorithm by which it is pos- sible to obtain the best subset of coefficients in order to improve results on phoneme recognition. Various strategies have been evaluated in order to improve the classifier initialization and the evolution itself. Classifi- cation results for a set of Spanish phonemes show the advantages of the propossed method and the speedup of the implemented parallelization. 1 Introduction Automatic speech recognition systems need a preprocessing stage to make the key features of the phonemes more evident, allowing to improve classification results [16]. This task is usually accomplished by different signal processing tech- niques like filter banks, linear prediction and cepstrum analysis [17]. Wavelet transform characteristics make this tool usefull for non stationary signal analysis [11]. Multiresolution analysis asociated with discrete wavelet transform can be implemented as a filter bank decomposition (or filter bank schemes) [23]. Wavelet packet transform (WPT) [6] is a generalization of the discrete wavelet transform (DWT) decomposition which offers a wider range of possibilities for signal representation. To compute this transform it is necesary to select a particular orthogonal basis among all the available basis (or filter banks). Nevertheless, in signal classification applications there is not evidential benefit on working with orthogonal basis. Without this restriction the result of the full WPT decomposition is a highly redundant set of coefficients, which is convenient to optimize in order to use it on signal classification. ? This work is supported by ANPCyT-UNER, under Project PICT No 11-12700, UNL- CAID 012-72 and CONICET. sinc( i) Laboratory for Signals and Computational Intelligence (http://fich.unl.edu.ar/sinc) L. D. Vignolo, D. H. Milone, H. L. Rufiner & E. M. Albornoz; "Parallel implementation for wavelet dictionary optimization applied to pattern recognition" 7th Argentine Symposium on Computing Technology. pp. 49-60, 2006.