Figure 9
From: Lip-Synching Using Speaker-Specific Articulation, Shape and Appearance Models
Mean correlations (together with standard deviations) between original and predicted trajectories for the main six articulatory parameters (jaw rotation, lip rounding, lower and upper lip opening, jaw retraction and larynx height). For each parameter, correlations for eleven conditions are displayed: the first correlation is for the trajectories predicted by concatenative synthesis using multirepresented diphones (see text); the second correlation is for trajectories predicted by HMM using acoustic boundaries; the rest of the data give results obtained after the successive iteration of the estimations of the phasing model. Asymptotic behavior is obtained within one or two iterations. From left to right: data from speakers CD, OC, and AA.