Machine learning models in ultrasound tongue imaging for the detection of children's speech disorders

点击翻译按钮获取中文摘要

Currently in Speech and Language Therapy (SLT), technological support is sparse. Assessment of speech disorders is particularly time consuming and suffers from a lack of technological solutions. Normally speech assessment involves listening to the child and writing down what they say. This approach can miss important subtleties in the way children speak. For example, a child may try to say "key" and it may be heard as "tea". This leads the SLT to believe the child cannot tell the difference between t and k and select a therapy designed to tackle this. However, using medical ultrasound to image tongue movements reveals that in many cases children are producing imperceptible errors. This is particularly the case for children with cleft lip and palate who produces a wide variety of unusual speech errors. However, ultrasound analysis is a time consuming task which can only be carried out by a speech scientist with specialist training. This project aims to use machine learning approaches for method for classifying tongue shapes in children with cleft lip and palate. This has the potential to be useful both for assessment and for measuring progress in intervention. Open source data from children with cleft lip and palate is available and an ongoing clinical project with NHS Greater Glasgow and Clyde is currently collecting further ultrasound data.