A machine learning ensemble to automatically classify tongue ultrasound contours based on displacement measurements

Simon Gonzalez

Wednesday, December 14th, 2022, 1.30pm – 2pm


This paper introduces a Machine Learning Ensemble Model to automatically classify tongue ultrasound contours. It has been trained on displacement measurements in English Coronal Obstruents (/t/, /s/, /tʃ/, /ʃ/), from eight female native speakers of Australian English. The model has an accuracy of 97.6% for the Random Forests and 74.4% for the Decision Tree. The accuracy is higher for fricatives than for stops. Results also show that the most reliable area for classification is from 20% to the 40% of the contour length, which corresponds to the tongue area between the tongue front and the tongue body.