Automatic discrimination of Polish stop consonants based on bursts analysis

The aim of the work reported is to test the possibility of speaker- and context-independent automatic discrimination of Polish stop consonants. A new approach to stop consonant discrimination has been proposed based on a linear combination of autocorrelation function values. By computing BETWEEN and WITHIN matrices for parameters representing different populations and by solving the general eigenproblem, the direction of the eigenvector is determined, corresponding to the maximum eigenvalue. When projected in this direction, objects belonging to one population are most clustered and pair-wise projection of microsegments ("each with each") representing stop consonants was performed. Multidimensional parameter space was reduced to one dimension (axis). The material consisted of nonsense words, with most common Polish stop consonant contexts, produced by 20 speakers (10 male and 10 female). Experiments were conducted for male and populations themselves maximally separated. It is in this direction that female voices separately as well as for all the voices pooled. The burst segment has been found to provide better cues for phone identification than the frication segment. The average identification rate (for voices pooled) was 764.