Classification of accurate and misarticulated /ɑr/ for ultrasound biofeedback using tongue part displacement trajectories

ABSTRACT

Ultrasound biofeedback therapy (UBT), which incorporates real-time imaging of tongue articulation, has demonstrated generally positive speech remediation outcomes for individuals with residual speech sound disorder (RSSD). However, UBT requires high attentional demands and may therefore benefit from a simplified display of articulation targets that are easily interpretable and can be compared to real-time articulation. Identifying such targets requires automatic quantification and analysis of movement features relevant to accurate speech production. Our image-analysis program TonguePART automatically quantifies tongue movement as tongue part displacement trajectories from midsagittal ultrasound videos of the tongue, with real-time capability. The present study uses such displacement trajectories to compare accurate and misarticulated American-English rhotic /ɑr/ productions from 40 children, with degree of accuracy determined by auditory perceptual ratings. To identify relevant features of accurate articulation, support vector machine (SVM) classifiers were trained and evaluated on several candidate data representations. Classification accuracy was up to 85%, indicating that quantification of tongue part displacement trajectories captured tongue articulation characteristics that distinguish accurate from misarticulated production of /ɑr/. Regression models for perceptual ratings were also compared. The simplest data representation that retained high predictive ability, demonstrated by high classification accuracy and strong correlation between observed and predicted ratings, was displacements at the midpoint of /r/ relative to /ɑ/ for the tongue dorsum and blade. This indicates that movements of the dorsum and blade are especially relevant to accurate production of /r/, suggesting that a predictive parameter and biofeedback target based on this data representation may be usable for simplified UBT.

KEYWORDS:

• Ultrasound
• tongue
• articulation errors
• rhotics
• child speech

Acknowledgments

The authors would like to thank Kathryn Eary, Michael Swearengen, Gregory Terrell, Sarah Stack, Maurice Lamb and Sarah Schwab for their contributions to this project, as well as all participants in the study. We are also grateful to Siemens Medical Solutions USA, Inc., for making the Acuson X300 scanner available for this project.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Notes

¹ Criterion for native speakers was exposure to an American English environment from 2 years of age onward. A few participants also had exposure to other languages in their households (including Japanese, Arabic or Mandarin Chinese) but were primarily in English dominant environments.

Additional information

Funding

This work was funded by the University of Cincinnati Creating Our Third Century funding support and National Institute on Deafness and Other Communication Disorders grants [R01 DC017301] (awarded to Suzanne Boyce, Michael A. Riley and T. Douglas Mast) and [R01 DC013668] (awarded to Suzanne Boyce and Doug Whalen).