A systematic review of the application of machine learning techniques to ultrasound tongue imaging analysis.

This article provides a comprehensive review of machine learning applications, particularly deep learning, in the analysis of ultrasound tongue image frame sequences (UTIFs) used to study tongue motion during speech production. It covers four main areas: foundational deep learning principles, motion tracking methodologies, feature extraction techniques, and cross-modality mapping between UTIFs and speech or other modalities. The review highlights the advantages of ultrasound imaging for real-time, non-invasive tongue visualization, discusses traditional and deep learning-based approaches for contour tracking and feature extraction, and examines articulatory-to-acoustic and acoustic-to-articulatory mapping methods relevant to silent speech interfaces and speech therapy. Challenges identified include limited annotated datasets, variability in tongue anatomy, noise in ultrasound images, and the need for real-time processing and clinical validation, with future directions emphasizing improved model generalization, data-efficient learning strategies, interdisciplinary collaboration, and the development of large foundation models for UTIF analysis.