Real-time speech MRI datasets with corresponding articulator ground-truth segmentations.

Area of Science:

• Medical Imaging
• Speech Science
• Machine Learning

Background:

• Real-time magnetic resonance imaging (rt-MRI) is increasingly used in speech science and clinical settings.
• Accurate segmentation of articulators and vocal tract is crucial for analyzing rt-MRI speech data.
• Existing rt-MRI speech datasets lack essential ground-truth segmentations for deep learning model development.

Purpose of the Study:

• To address the barrier of missing ground-truth data for rt-MRI speech analysis.
• To present novel rt-MRI speech datasets with manual ground-truth segmentations.
• To provide accessible code for state-of-the-art deep learning segmentation methods.

Main Methods:

• Acquired rt-MRI speech data from five healthy adult volunteers using standard clinical MRI equipment.
• Manually created ground-truth segmentations for six key anatomical features (tongue, soft palate, vocal tract, etc.).
• Developed and documented code for a current state-of-the-art deep learning segmentation method.

Main Results:

• Successfully generated comprehensive rt-MRI speech datasets with accurate ground-truth segmentations.
• Included velopharyngeal closure patterns in the dataset.
• Publicly released the datasets, code, and implementation instructions.

Conclusions:

• The presented datasets and code provide a foundational resource for the speech science and machine learning communities.
• Facilitates the development and validation of deep learning models for rt-MRI speech segmentation.
• Enables further research into speech articulation and related clinical applications.