Contributions to a Quantitative Unsupervised Processing and Analysis of Tongue in Ultrasound Images

摘要

Speech production studies and the knowledge they bring forward are of paramount importance to advance a wide range of areas including Phonetics, speech therapy, synthesis and interaction. Several technologies have been considered to study static and dynamic features of the articulators and speech motor control, such as electromagnetic articulography (EMA), real-time magnetic resonance (RTMRI) and ultrasound (US) imaging. While the latest advances in RTMRI provide a wealth of data of the full vocal tract, it is an expensive resource that requires specialized facilities. In this sense, US is a more affordable alternative for several contexts, enabling the acquisition of larger datasets, but demands adequate computational approaches for processing and analysis. While the literature is prolific in proposing methods for tongue segmentation from US, the noisy nature of the images and the specificities of the equipment often dictate a poor performance on novel datasets, a matter that needs to be assessed, before large data acquisition, to devise suitable acquisition and processing methods. In the scope of a line of research studying speech changes with age, this work describes the first results of an automatic tongue segmentation method from US, along with a characterization of the main challenges posed by the image data. Even though improvements are still needed, particularly to ensure temporal coherence, at its current stage, this method can already provide the required data for an automatic analysis of maximum tongue height, a relevant parameter to assess speech changes on vowel production.