Synthesis of sentence-level speech based on measured vocal tract area functions

A parametric model of the vocaf tract is developed based on an inventory of area functions ~quired for one male subject with Magnetic Resonance Imaging (MRI). The model is then used to synthesize a sentence recorded from the same sub ject. AREA FUNCTION MODEL OF THE VOCAL TRACT The question that is addressed in this paper is whether an inventory of vocal tract area functions, which represent only stat ic speech sounds, can be successfully used to synthesize dynamic (i.e. connected) speech recorded from a speaker. This is a form of articulator synthesis in which the effective shape of vocal tract airway is directly manipulated rather than explicitly defining position and movement of individual articulators (e.g. tongue tip and body, velure, lips, jaw, etc.). For this discussion, an area function inventory for one male speaker acquired with Magnetic Resonance Imaging (MRI) (1) will be used. The area function at any instant of time is represented as the combination of a vowel substrate and an imposed consonantal element. Such an approach ww proposed by (2) and further developed by (3), and (4). The vowel substrate V for this study is represented by a principal components analysis (PCA~ of the ten vowel area functions in (1,5) as, v(z) = &cj#j(z) +n(z) (1) j=l where z the distance from the glottis, the dj (z)’s are the principal components (or eigenvectors), the cj’s are the coefficients produced by the PCA that will reconstruct each of the original ten vowels, and Q(z) is the mean area function. Thus, any vowel is represented as a perturbation around the mean area function. It has been found that four of the principal components (i.e. 4 coefficients) can reconstruct each area function to within 3 percent of its original fidelity. For a time-varying area function, the cj’s become time-dependent parameters, N V(z,t) = ~cj(t)dj(z) + Q(x) . (2)