Vocal Tract Visualization
The goal of vocal tract visualization is to allow us to view the vocal
tract during unimpeded speech production. Knowledge about the vocal
tract is useful for clinical diagnosis of speech disorders, for
providing more accurate data for acoustic modeling of speech
production, and for studying the mechanics of vocal tract.
Tongue Surface Reconstruction from Ultrasound Data
Ultrasound data collection is well suited to the vocal tract for
several reasons. Ultrasound is non-invasive. We can collect
tongue surface scans with a transducer held under the chin. With
a compressible standoff, the tongue and jaw are free to make
normal speech movements. Ultrasound is safe. It uses high
frequency (~5mhz) sound which pose no danger to the subjects
in repeated trials. Ultrasound is fast. Ultrasound slices can be
collected at video rates (30fps).
Ultrasound Data
Ultrasound coronal (side to side) scan of tongue surface during speech .
Using a developmental 3D ultrasound transducer, we collect 60 coronal
slices, oriented radially in space. The 3D transducer has a standard
128 crystal array (capable of collecting a single 2D slice) mounted
on a motorized pivot, allowing it to rotate in the third dimension.
It collects up to 60 slices, each 1 degree apart. This process takes
about 10 seconds.
The upper white line in the above image is caused by the density change
at the tissue/air interface at the surface of the tongue. Note the deep
groove in the middle of the tongue (this is during the production of
the sound "e").
Image Processing
Detection of the tongue surface in each ultrasound slice was done using
a dynamic programming contour tracking algorithm [1].
This is done for each of the 2D coronal slices.
Surface Reconstruction
Reconstruction of the surface is simplified by the automated collection
procedure. The rotation axis of the transducer is known, as is the
rotation between slices. The 2D point sets for the tongue surface in
each slice are returned to their relative 3D coordinates, and an
interpolating bspline surface is fitted to the points.
Visualization
Reconstruction of the tongue during production on an "e".
The reconstructed tongue surface can be visualized to provide an intuitive
grasp of the tongue shape. It can also be statistically measured for more
quantitative analysis.
Research Concerns
Our current methods include detection of the 2D tongue surface in each
ultrasound slice. I am currently seeking a method to use the neighboring
slice data in a 3D surface detection process. Other future work includes
doing reconstructions from multiple pass 2D collections of continuous
speech.
References
[1] Michael Unser, and Maureen Stone. Automated detection of the tongue surface
in sequences of ultrasound images. J. Acoust. Soc. Am.,(5),
May 1992:3001-3007.
lundberg@speech.umaryland.edu
Jump to Andrew's Home Page