First International Conference on Spoken Language Processing (ICSLP 90)
This paper presents two models of contextual effects which can cope with co-articulation problems, especially vowel neutralization. Model 1 predicts target spectral peaks in reduced vowels based on interactions between spectral peak pairs. Model 2 distorts a spectral pattern space to pull back reduced spectral patterns into each of their correct categories based on influences of preceding spectral patterns on phoneme boundaries in a spectral pattern space. To construct two models and to substantiate these models, two psychoacoustic experiments were carried out which measured the extent of phoneme boundary shift with (1) a single formant stimulus as a preceding anchor and (2) a vowel as a preceding anchor. The results from experiment 1 showed that the contextual effect between single formant stimuli should play an important role in phoneme neutralization recovery, and that the neutralization recovery model is formulated as the sum of the contextual effects resulting from interaction between spectral peaks. Additionally, a comparison of the results of the first and second experiments showed that a phoneme boundary shift with a vowel anchor can be considered the sum of a shift with a single formant anchor and a factor from the preceding stimulus. This factor was represented as a function of the distance between the preceding vowel anchor and the perceived vowel in a phoneme space.
Bibliographic reference. Akagi, Masato (1990): "Contextual effect models and psycho acoustic evidence for the models", In ICSLP-1990, 569-572.