Articulatory speaker normalisation based on MRI-data using three-way linear decomposition methods

The aim of this study was to characterise, to model and to compare the different lingual articulatory strategies of a group of speakers. Individual principal component analysis (PCA) and multi-linear decomposition methods have been applied to different representations of the tongue contour extracted from magnetic resonance images (MRI). The corpus consisted of seven speakers articulating 63 French vowels and consonants. On the average, over the seven speakers, the Root Mean Square prediction Error (RMSE) was 0.12 cm accounting for a percentage of variance explanation of 92.6% for the individual PCA, using 4 components. Several Multi-linear decomposition methods, to model the tongue contour with a single set of components, have been performed and compared. The 2-LevelPCA gave the best results among the other techniques. By means of a Student's t-test, at 5% of significance level, we found that 2level-PCA equals the PCA performance with 11 components to represent 91% of the variance explanation with a RMSE of 0.11 cm. While the same method, with 4 components, represents 75% of the variance explanation with a RMSE of 0.19 cm.