Speaker normalization of stressed and unstressed vowels in articulatory and formant spaces

By C. Geng and C. Mooshammer

 

One of the goals of phonetic investigations is to find strategies for vowel production independent of speaker-specific vocal-tract anatomies and individual biomechanic proporerties. In this paper we apply a technique for speaker normalization to formant spaces, lingual vowel target positions and palate midlines which in the literature sometimes is referred to as Procrustes Analysis. Procrustes Analysis comes in two guises: One, the orthogonal version, is based on Euclidean geometry. This means that only transformations are allowed which preserve the angles between the measured landmarks or objects. It is more apt for the analysis and normalization of formant spaces. In contrast, the second version of the model is based on affine geometry, which means that the angles between corresponding landmarks do not have to be preserved. It has its strengths in the articulatory domain. We can show the relevance of these models by (a) analysing the remaining variances after application of these procedures to articulatory (vowel targets in the mid of the acoustically defined vowel) as well as acoustic (related formants) data and by (b) relating the transformations applied to the articulatory target configurations to anatomical landmarks (EPG midline traces).