Barton F. Lane, Eric Vatikiotis-Bateson
Reconstructing the Vocal Tract During Vowel Production using Magnetic Resonance Images
Abstract:The goal of this study was to reconstruct and accurately quantify a three-dimensional (3D) vocal tract during the pronunciation of different vowels. This was done using magnetic resonance image (MRI) data recorded during sustained pronunciation of five vowels. Since MRI cannot distinguish between bone and air, it has been difficult to reconstruct reliable volumes for the oral cavity since the teeth do not produce an air-tissue boundary. Therefore, a major task in this study was to find a means to identify the teeth in the images and subsequently to subtract them from the reconstructed oral cavity volume. Of the five vowels examined, two had cavity shapes suitable for tooth subtraction using threshold-detection techniques. Cross-validation of the subtracted volumes with a more direct, physical measure of tooth volume suggest an accuracy of more than 90 percent. Using configurations in which accurate subtraction is possible to derive constant values for the teeth it is then possible to apply these values and to estimate reasonable volumes for all configurations produced by the same speaker.