CONICET | Buscador de Institutos y Recursos Humanos

Many computational theories have been developed to improve artificial phonetic classificationperformance from linguistic auditory streams. However, less attention has been given topsycholinguistic data and neurophysiological features recently found in cortical tissue. Wefocus on a context in which basic linguistic units?such as phonemes?are extracted androbustly classified by humans and other animals from complex acoustic streams in speechdata. We are especially motivated by the fact that 8-month-old human infants can accomplishsegmentation of words from fluent audio streams based exclusively on the statisticalrelationships between neighboring speech sounds without any kind of supervision. In thispaper, we introduce a biologically inspired and fully unsupervised neurocomputationalapproach that incorporates key neurophysiological and anatomical cortical properties,including columnar organization, spontaneous micro-columnar formation, adaptation to contextualactivations and Sparse Distributed Representations (SDRs) produced by means ofpartial N-Methyl-D-aspartic acid (NMDA) depolarization. Its feature abstraction capabilitiesshow promising phonetic invariance and generalization attributes. Our model improves theperformance of a Support Vector Machine (SVM) classifier for monosyllabic, disyllabic andtrisyllabic word classification tasks in the presence of environmental disturbances such aswhite noise, reverberation, and pitch and voice variations. Furthermore, our approachemphasizes potential self-organizing cortical principles achieving improvement without anykind of optimization guidance which could minimize hypothetical loss functions by meansof?for example?backpropagation. Thus, our computational model outperforms multiresolutionspectro-temporal auditory feature representations using only the statistical sequentialstructure immerse in the phonotactic rules of the input stream.