Modeling of contextal effect based on spectral peak interaction

Abstract

This paper presents a model of a lower level contextual effect that can cope with coarticulation problems, especially vowel neutralization. The model is constructed to overshoot spectral peak trajectories based on spectral peak interaction, assuming that the lower level contextual effect is represented as the sum of interaction between each spectral peak pair. The interaction function is determined experimentally in order to reduce the distance between a real spectral peak and its target which is a spectral peak mean computed for vowel uttered in isolation. The interaction function thus determined suggests that: (1) there can be a time-frequency lateral inhibition in the auditory system like that on the retina in the visual system, (2) the interaction function is consistent with the results of psychaocoustic experiments concerning the assimilation and/or contrast effect using paired single formant stimuli, and (3) the contextual effect between adjacent phonemes can be represented as the sum of the assimilation and/or contrast effects between each spectral peak pair. Applying the determined interaction function to real speech data to cope with coarticulation problem, spectral peak trajectories overshoot, spectral peaks at the vowel center approach their own targets, and the distance between each vowel category pair increases.