Learning and adaptation of a tongue shape modelwith missing data

摘要

Using data-driven techniques and ultrasound data, it is possible to learn models that reconstruct the tongue shape of a speaker with submillimetric accuracy given the location of 3–4 fleshpoints, and to adapt these models to a new speaker for which little data is available. In practice, tongue contours extracted from ultrasound imaging are often incomplete because of shadowing, noise and other factors. We extend these models to deal with missing data during learning and adaptation, and show that submillimetric accuracy can still be achieved even with relatively large amounts of missing data.