Automatic tracking of 3D vocal tract features during speech production using MRI

Abstract

Magnetic resonance imaging has many advantages for visualising the process of speech production, but an important disadvantage is the long scanning acquisition time relative to the characteristic time of articulator motion of tenth of second. Southampton Dynamic Magnetic Resonance Imaging is a technique developed in a previous project to solve this problem. This technique achieves an apparent high temporal resolution suitable for dynamic studies. Consequently, a large number of images can be generated describing the evolution of the vocal tract shape. This makes a manual extraction of the vocal tract shape a tedious and time consuming process. The aim of this project firstly is to improve and extent the SDMRI method, and secondly, to determine the outline of the vocal tract automatically. Different features extraction techniques were analysed and two of them were combined to make a new automatic shape extraction tool, i.e. the active shape models and the Hough transform. Active shape models describe the shape of the articulators while the Hough transform locates it with no initialisation. Initially, the new algorithm was tested analysing isolated magnetic resonance images for extracting tongue shapes; however, although the results were satisfactory the algorithm often fails when multiple solutions are present. A global analysis of the image sequence overcomes these difficulties and the dynamic Hough transform was adapted for our purposes. Experimental results reveal that the algorithm does indeed find the correct shape and position of the tongue and also that it is robust under noisy conditions. The model was extended to other articulators, i.e. the lips. This approach leads to a new algorithm for automatic extraction of articulatory shape in magnetic resonance image sequences as evident in the results presented in this thesis.

More information

Identifiers

Catalogue record

Export record

ASCII CitationAtomBibTeXData Cite XMLDublin CoreDublin CoreEP3 XMLEndNoteHTML CitationHTML CitationHTML ListJSONMETSMODSMPEG-21 DIDLOpenURL ContextObjectOpenURL ContextObject in SpanRDF+N-TriplesRDF+N3RDF+XMLRIOXX2 XMLReferReference ManagerSimple Metadata