In this paper we consider a feature estimation approach for vocal fold pathology classification, based on digital signal processing theory. This problem is addressed by formulating a stochastic maximum likelihood (ML) estimation procedure, based on Estimation-Maximization (EM) algorithm. New spectral parameters of speech, noted as Spectral Pathology Component (SPC) is estimated. For classification purposes, the counterpropagation neural network (CNN) was proposed. A set of log Mel-frequency filter bank coefficients were used to parametrize the SPC spectral feature. An evaluation of CNN based classifier were performed using speech recording from healthy and pathology patients.