Deep Learning in Ultrasound Tongue Imaging: A Systematic Review Toward Automated Detection of Speech Sound Disorders

Saja Al Ani^1*Joanne Cleland²Ahmed Zoha¹

• ¹University of Glasgow, Glasgow, United Kingdom
• ²University of Strathclyde, Glasgow, United Kingdom

Background: Speech sound disorders (SSD) in children can significantly impact communication and development. Ultrasound tongue imaging (UTI) is a non-invasive method for visualising tongue motion during speech, offering a promising alternative for diagnosis and therapy. Deep learning (DL) techniques have shown great promise in automating the analysis of UTI data, although their clinical application for SSD remains underexplored. Objective: This review aims to synthesise how DL has been utilised in UTI to support automated SSD detection, highlighting the advancement of techniques, key challenges, and future directions. Methods: A comprehensive search of IEEE Xplore, PubMed, ScienceDirect, Scopus, Taylor & Francis, and arXiv identified studies from 2010 through 2025. Inclusion criteria focused on studies using DL to analyse UTI data with relevance to SSD classification, feature extraction, or speech assessment. ElevenEight studies met the criteria: three directly tackled disordered speech classification tasks, while four addressed supporting tasks like tongue contour segmentation and tongue motion modelling. Promising results were reported in each category, but limitations such as small datasets, inconsistent evaluation, and limited generalisability were common. Results: DL models demonstrate effectiveness in analysing UTI for articulatory assessment and show early potential in identifying SSD-related patterns. The included studies collectively outline a developmental pipeline, from foundational pre-processing to phoneme-level classification in typically developing speakers, and finally to preliminary attempts at classifying speech errors in children with SSD. This progression illustrates significant technological advances; however, it also emphasises gaps such as the lack of large, disorder-focused datasets and the need for integrated end-to-end systems. Conclusions: The field of DL-driven UTI assessment for speech disorders is developing. Current studies provide a strong technical foundation and proof-of-concept for automatic SSD detection using ultrasound, but clinical translation remains limited. Future research should prioritise the creation of larger annotated UTI datasets of disordered speech, developing generalisable and interpretable models, and validating fully integrated DL-UTI pipelines in real-world speech therapy settings. With these advances, DL-based UTI systems have the potential to transform SSD diagnosis and treatment by providing objective, real-time articulatory feedback in a child-friendly manner.