Classificação de Voz versus Silêncio via Dicionários Redundantes

This paper presents a technique to solve the pro- blem of classifying portions of signal between voice and silence. Surprisingly, the proposed technique is not based on the variation of energy levels over the signal, but on the fundamental charac- teristics that determine the essence of each of these two classes of signals, namely voice and silence. To this end, a redundant dictionary of basis functions (atoms) is built and the signal is analyzed via Matching Pursuit. From this analysis, the technique training phase, the a priori discrete probability distribution of occurrence of a set of atoms for each class of interest is computed, allowing subsequent discrimination between the classes. The paper presents promising results for signals with high signal to noise ratio.