Modeling the German Stress Distinction

Low dimensional and speaker-independent linear vocal tract parametrizations can be obtained using the 3-mode PARAFAC factor analysis procedure first introduced by Harshman et al. (1977). The following study used PARAFAC to investigate the stress distinction in German vowel production. Tongue movements of six German speakers were recorded by means of EMMA. The speech material consisted of the 15 German vowels, recorded in /t/context. Our corpus includes these vowels in stressed and unstressed position. They were entered into the classical PARAFAC1 model treating the stress distinction for each subject as two different speakers. This gave a reasonable 2-factor solution, but was not without drawbacks. The model turned out to be capable of recovering gross anatomical properties of our subjects, but failed to return intraindividual differences in tongue shapes with respect to word stress. This indicated that the strict linearity assumptions required in the classical PARAFAC model were too strong to capture stress-specific variation in full detail. We supposed that a model closely related to PARAFAC, PARAFAC2, should allow to account for systematic variation produced by word stress by imposing weaker structure on the data. As will be shown, PARAFAC2 modeled the physical properties of the vocal tract shape in a more realistic and plausible way.