Realce de Sinais de Voz em Presença de Ruídos Acústicos Não-Estacionários utilizando o Método EMD

This paper presents a speech enhancement technique based on the empirical mode decomposition method. The EMD-based filtering is evaluated in speech utterances corrupted with six acoustic noises and six different signal-to-noise ratios. The noises present different values of INS (index of nonstationarity). For comparison, another speech enhancement technique, based on the use of IMCRA and OMLSA estimators, is used in the experiments. The comparison among the techniques considers two different measures: the segmental signal-to-noise ratio and the accuracies in a speaker identification task. The results show that the enhancement performance is strongly dependent on the INS of the noises.