TRACTUS

TRACTUS software

TRACTUS (Temporally Resolved Articulatory Configuration Tracking of UltraSound) is a suite of Matlab (Mathworks, 2014) functions designed to perform temporal articulatory analysis on ultrasound images without the need for tracing of tongue contours. This software suite is freely available for download at the website address below (Carignan, 2014).

TRACTUS uses principal component analysis (PCA) to identify a set of linearly uncorrelated variables (principal components, PCs) which account for the greatest amount of image variance in the data set. The software generates PC scores for each ultrasound frame and maps PC loadings onto their original spatial orientation (PC heatmaps). The PC scores for each ultrasound frame can be plotted sequentially and analyzed as temporal vectors which determine the extent to which each ultrasound frame is correlated with the given PC and associated heatmap.

TRACTUS demo #1: vowel articulation

An example is given below of how TRACTUS can be used to generate signals for vowel articulation, demonstrated here with lingual tensing of the word "ban" produced by a speaker of American English. The lingual tensing signal was created by transforming the articulatory PCs to correlate with the front diagonal of the vowel space via linear regression. Clockwise from the top left, these images show:

1. The raw ultrasound frames.
2. The lingual tensing signal generated using TRACTUS and the corresponding spectrogram.
3. PC loading heatmap associated with the lingual tensing dimension.
4. Composite frames generated from the cumulative weightings of the first 20 PCs; these frames approximate the original raw ultrasound images but are generated solely from the PC loadings and scores for each ultrasound image. They have also been rotated counter-clockwise to be parallel to the speaker's occlusal plane.

TRACTUS demo #2: consonant articulation

An example is given below of how TRACTUS can be used to generate signals for consonant articulation, as well. These signals were created by including the 20 PC scores as inputs to separate linear discriminant analysis (LDA) models trained on ultrasound frames associated with velar and alveolar consonant productions. The signals demonstrate the usability of this technique for researching temporal aspects of co-articulation, for example.

References

Carignan, C. (2014). TRACTUS (Temporally Resolved Articulatory Configuration Tracking of UltraSound) software suite. URL: http://christophercarignan.github.io/TRACTUS.

Mathworks (2014). MATLAB R2014a (Version 8.3). URL: http://www.mathworks.com/products/matlab.