A multispeaker dataset of raw and reconstructed speech production real-time MRI video and 3D volumetric images

Real-time magnetic resonance imaging (RT-MRI) of human speech production is enabling significant advances in speech science, linguistics, bio-inspired speech technology development, and clinical applications. Easy access to RT-MRI is however limited, and comprehensive datasets with broad access are needed to catalyze research across numerous domains. The imaging of the rapidly moving articulators and dynamic airway shaping during speech demands high spatio-temporal resolution and robust reconstruction methods. Further, while reconstructed images have been published, to-date there is no open dataset providing raw multi-coil RT-MRI data from an optimized speech production experimental setup. Such datasets could enable new and improved methods for dynamic image reconstruction, artifact correction, feature extraction, and direct extraction of linguistically-relevant biomarkers. The present dataset offers a unique corpus of 2D sagittal-view RT-MRI videos along with synchronized audio for 75 participants performing linguistically motivated speech tasks, alongside the corresponding public domain raw RT-MRI data. The dataset also includes 3D volumetric vocal tract MRI during sustained speech sounds and high-resolution static anatomical T2-weighted upper airway MRI for each participant. Measurement(s) Speech • vocal tract Technology Type(s) magnetic resonance imaging Factor Type(s) age • sex • native/non-native speaker Sample Characteristic - Organism Homo sapiens Measurement(s) Speech • vocal tract Technology Type(s) magnetic resonance imaging Factor Type(s) age • sex • native/non-native speaker Sample Characteristic - Organism Homo sapiens Machine-accessible metadata file describing the reported data: https://doi.org/10.6084/m9.figshare.14892921

Stimuli and linguistic justification. The stimuli were designed to efficiently capture salient, static and dynamic, and articulatory and morphological aspects of speech production of American English in a single 90-minute scan session. Table 2 lists the speech stimuli used for the RT-MRI data collection of the first 45-minute sub scan session. Each individual task was designed to be performed within 30 seconds at a normal speaking rate, however the actual recordings varied in duration depending on the length of the task and the natural speaking rate of the individual. The stimulus set contained material to elicit both scripted speech and spontaneous speech. The scripted speech tasks were consonant production in symmetric vowel-consonant-vowel context, vowels /V/ produced between the consonants /b/ and /t/, i.e., in /bVt/ contexts, four phonologically rich sentences 35 , and three reading passages commonly used in speech evaluation and linguistic studies [36][37][38] . Scripted instructions to produce several gestures were also included. These gestures were clenching, wide opening of mouth, yawning, swallowing, slow production of the sequence /i/-/a/-/u/-/i/, tracing of the palate with the tongue tip, and singing "la" at their highest and lowest pitches. The stimuli in the scripted speech were repeated twice. The spontaneous speech tasks were to describe the content and context of five photographs (Supplemental Fig. 1) and to answer five open-ended questions (e.g., "What is your favorite music?"). The full scripts and questions are provided in Table 3. Table 4 lists the speech stimuli used for the 3D volumetric data collection of a 30-minute part of the scan session. Each individual task was designed for a participant to sustain vowels, consonant sounds, or to maintain www.nature.com/scientificdata www.nature.com/scientificdata/ postures for 7 seconds. The stimulus set contained consonant production in symmetric vowel-consonant-vowel context, vowels /V/ produced between the consonants /b/ (or /p/) and /t/, and production of several postures. All of the stimuli were repeated twice.
RT-MRI acquisition. RT-MRI acquisition was performed using a 13-interleaf spiral-out spoiled gradient-echo pulse sequence 33 . This is an efficient scheme for sampling MR measurements, each with the different initial angle being interleaved by the bit-reversed order in time 39 . The 13 spiral interleaves, when collected together, fulfil the Nyquist sampling rate. Imaging parameters were: repetition time = 6.004 ms, echo time = 0.8 ms, field-of-view (FOV) = 200 × 200 mm 2 , slice thickness = 6 mm, spatial resolution = 2.4 × 2.4 mm 2 (84 × 84 pixels), receiver bandwidth = ± 125 kHz, flip angle = 15°. Imaging was performed in the mid-sagittal plane, which was prescribed using a real-time interactive imaging platform 40 (RT-Hawk, Heart Vista, Los Altos, CA). Real-time visualization was implemented within the custom platform by using a sliding window gridding reconstruction 41 to ensure participant's compliance with stimuli and to detect substantial head movement.
Acquisition was divided into 20-40 second task intervals presented in Table 2, each followed by a pause of the same duration so as to allow enough a brief break for the participant prior to the next task and to avoid gradient heating.
RT-MRI Reconstruction. The dataset provides one specific image reconstruction that has been widely used in speech production research 26,33 . This image reconstruction involves optimizing the following cost function: where A represents the encoding matrix that models for the non-uniform fast Fourier transform and the coil sensitivity encoding, m is the dynamic image time series to be reconstructed, d is the acquired multiple-coil raw data, λ is the regularization parameter, ∇ t is the temporal finite difference operator, and ⋅ 2 2 and ⋅ 1 are l 2 and l 1 norms, respectively. The coil sensitivity maps were estimated using the Walsh method 42 from temporally combined coil images. The regularization term encourages voxels to be piecewise constant along time. This regularization has been successfully applied to speech RT-MRI 26,43,44 and a variety of other dynamic MRI applications where the primary features of interest are moving tissue boundaries [45][46][47] .
Our reference implementation solves Equation [1] using nonlinear conjugate gradient algorithm with Fletcher-Reeves updates and backtracking line search 48 . The algorithm was terminated either at 150 maximum iterations or the step size fell below <1e-5 during line search. Reference images were reconstructed using 2 spiral arms per time frame, resulting in a temporal resolution of 83.28 frames per second. This temporal resolution is enough to capture important articulator motions 1 , but reconstruction at different temporal resolutions is also possible with the provided dataset by adjusting the number of spiral arms per time frame. The algorithm was implemented in both MATLAB (The MathWorks, Inc., Natick, MA) and Python (Python Software Foundation, https:// www.python.org/). The provided image reconstruction was performed in MATLAB 2019b on a Xeon E5-2640 v4 2.4 GHz CPU (Intel, Santa Clara, CA) and a Tesla P100 GPU (Nvidia, Santa Clara, CA). Reconstruction time was 160.69 ± 1.56 ms per frame. Parameter selection of λ is described in the technical validation section below.
Audio data. One main microphone was positioned ~0.5 inch away from the participant's mouth. The microphone signal was sampled at 100 kHz each. The data were recorded on a laptop computer using a National Instruments NI-DAQ 6036E PCMCIA card. The audio sample clock was hardware-synchronized to the MRI scanner's 10 MHz master clock. The audio recording was started and stopped using the trigger pulse signal from the scanner. The real-time audio data acquisition routine was written in MATLAB. Audio was first low-pass filtered and decimated to a sampling frequency of 20 kHz. The recorded audio was then enhanced using a normalized least-mean-square noise cancellation method 34 and was aligned with the reconstructed MRI video sequence to aid linguistic analysis.
3D Volumetric MRI of sustained sounds. An accelerated 3D gradient-echo sequence with Cartesian sparse sampling was implemented to provide static high-resolution images of the full vocal tract during sustained sounds or postures. Imaging parameters were: repetition time = 3.8 ms, spatial resolution = 1.5625 × 1.5625 × 1.5625 mm 3 , FOV = 250 × 250 × 125 mm 3 (respectively in the anterior-posterior,  www.nature.com/scientificdata www.nature.com/scientificdata/ superior-inferior, and left-right directions), image matrix size = 160 × 160 × 80, and flip angle = 5°. The central portion (40 × 20) of the k y -k z space was fully sampled to estimate the coil sensitivities from the data itself. The outer portion of the k y -k z space was sampled using a sparse Poisson-Disc sampling pattern, which together resulted in 7-fold net acceleration, and a total scan time of 7 seconds.
Data were acquired while the participants sustained for 7 seconds a sound from the full set of American English vowels and continuant consonants ( Table 4). The stimuli were presented to the participant via a projector-mirror setup, and upon hearing a "GO" signal given by the scanner operator, a participant started to sustain the sound or posture; this was followed by the operator triggering the acquisition manually. Participants sustained the postures as long as they could hear the scanner operating. A recovery time of 5-10 seconds was given to the participant between the stimuli.
Image reconstruction was performed off-line by a sparse-SENSE constrained reconstruction 49 . This method reconstructs images from multichannel, undersampled data, similar to the optimization problem shown in Equation [1]. One difference is that the non-Cartesian Fourier encoding in matrix A is replaced by undersampled Cartesian Fourier encoding. Also the temporal finite difference sparsity constraint is replaced by an isotropic spatial total variation 33,50 . The reconstruction was achieved using the open-source Berkeley Advanced Reconstruction Toolbox (BART) 51 ; the image reconstruction was performed in MATLAB using GPU acceleration. T2-Weighted MRI at rest. Fast spin echo-based sequence was performed to provide high-resolution T2-weighted images with fine detail anatomical characteristics of the vocal tract at rest position. Imaging was run to obtain full sweeps of the vocal tract in the axial, coronal, and sagittal orientations, for each of which the number of slices ranges from 29 to 70, depending on the size of the vocal tract. Imaging parameters were: repetition time = 4600 ms, echo time = 120-122 ms, slice thickness = 3 mm, in-plane FOV = 300 × 300 mm 2 , in-plane spatial resolution = 0.5859 × 0.5859 mm 2 , number of averages = 1, echo train length = 25, and scan time = approximately 3.5 minutes per orientation.

Data Records
This dataset is publicly available in figshare 52 . The total size of this dataset is approximately 966 GB. It contains (i) raw 2D sagittal-view RT-MRI data, reconstructed images and videos, and synchronized denoised audio, (ii) 3D volumetric MRI images, and (iii) T2-weighted MRI images. The data for participant (subject) XYZ is contained in folder with identifier subXYZ (e.g., sub001/) and organized into three main folders: 2D RT-MRI data are located in the subfolder 2drt/ (e.g., sub001/2drt/), 3D volumetric images in 3d/, and T2-weighted images in t2w/. The contents and data structures of the dataset are detailed as follows.

Stimulus Index
Script or question/topic vcv1 apa upu ipi ata utu iti aka uku iki aba ubu ibi ada udu idi aga ugu igi vcv2 atha uthu ithi asa usu isi asha ushu ishi ama umu imi ana unu ini ala ulu ili vcv3 afa ufu ifi ava uvu ivi ara uru iri aha uhu ihi awa uwu iwi aya uyu iyi bvt beet bit bait bet bat pot but bought boat boot put bite beaut bird boyd abbot shibboleth She had your dark suit in greasy wash water all year. Don't ask me to carry an oily rag like that. The girl was thirsty and drank some juice followed by a coke. Your good pants look great however your ripped pants look like a cheap version of a k-mart special is that an oil stain on them.
rainbow When the sunlight strikes raindrops on the air they act as a prism and form a rainbow. The rainbow is a division of white light into many beautiful colors. These take the shape of a long round arc with its path high above and its two ends apparently beyond the horizon. There is according to legend a boiling pot of gold at one end. People look but no one ever finds it. When a man looks for something beyond his reach his friends say he is looking for the pot of gold at the end of the rainbow. grandfather1 You wish to know all about my grandfather. Well, he is nearly ninety three years old yet he still thinks as swiftly as ever.
He dresses himself in an old black frock coat usually several buttons missing. A long beard clings to his chin, giving those who observe him a pronounced feeling of the utmost respect. When he speaks, his voice is just a bit cracked and quivers a bit. grandfather2 Twice each day he plays skillfully and with zest upon a small organ. Except in the winter when the snow or ice prevents, he slowly takes a short walk in the open air each day. We have often urged him to walk more and smoke less but he always answers, "banana oil" grandfather likes to be modern in his language.

northwind1
The north wind and sun were disputing which was the stronger, when a traveler came along wrapped in a warm cloak. They agreed that the one who first succeeded in making the traveler take off his cloak should be considered stronger than the other. northwind2 Then the north wind blew as hard as he could, but the more he blew the more closely did the traveler fold his cloak around him and at last the north wind gave up the attempt. Then the sun shone out warmly and immediately the traveler took off his cloak and so the north wind was obliged to confess that the sun was stronger of the two.  www.nature.com/scientificdata www.nature.com/scientificdata/ RT-MRI. Raw RT-MRI data are provided in the vendor-agnostic MRD format (previously known as ISMRMRD, https://ismrmrd.github.io/) 53 , which stores k-space MRI measurements, k-space location tables, and sampling density compensation weights. Parameters for the acquisition sequence are contained in the file header information. In addition, this dataset includes reconstructed image data for each participant and task in HDF5 format, audio files in WAV format, and videos in MPEG-4 format.
For each participant, RT-MRI raw data is contained in the subfolder raw/ (e.g., sub001/2drt/raw/), reconstructed image data in recon/, co-recorded audio (after noise cancellation) in audio/, and reconstructed speech videos with aligned audio in video/. Table 5 summarizes the data structure and naming conventions for this dataset. The imaging DICOM files were de-identified using the Clinical Trial Processor (CTP) developed by the Radiological Society of North America (RSNA) 54 . Specifically, data anonymization was completed using a command line tool developed in the Java programming language (http://mircwiki.rsna.org/index.  www.nature.com/scientificdata www.nature.com/scientificdata/ php?title=The_DicomAnonymizerTool). All images followed the standardized DICOM format and some of the attributes were removed or modified to preserve privacy of the participants, specifically: PatientName was modified to follow the subXYZ pattern, and the original study dates were shifted by the same offset for all participants.

Metadata.
We provide presentation slides that contain the experimental stimuli including scripts and pictures used for the visualization to the participants, in PPT format in Stimuli.ppt. Demographic information for each participant is contained in XLSX format in Subjects.xlsx. This meta file contains sex, race, age (at the time of scan), height (cm), weight (kg), origin, birthplace, cities raised and lived, L1 (first language), L2 (second language, if any), L3 (third language, if any), and the first language and birthplace of each participant's parents.
Additionally, meta information for each participant and RT-MRI task is contained in JSON format in metafile_ public_ < timestamp > .json. For each participant, we include the following demographic information: 1) L1 (first language), 2) L2 (second language, if any), 3) sex (M/F), 4) age (years at the time of scan). Further, visual and audio quality assessment scores are provided using a 5-level Likert scale for each participant stratified into categories: off-resonance blurring (1, very severe; 2, severe; 3, moderate; 4, mild; 5, none), video SNR (1, poor; 2, fair; 3, good; 4, very good; 5, excellent), aliasing artifacts (1, very severe; 2, severe; 3, moderate; 4, mild; 5, none), and audio SNR (1, poor; 2, fair; 3, good; 4, very good; 5, excellent). Specifically, an MRI expert with 6 years of experience in reconstructing and reading speech RT-MRI examined all the RT-MRI videos reconstructed for each participant and assessed and scored their visual and audio quality. For each task of the individual participants, the information about task's index, name, existence of file, and notes taken during data inspection by inspectors are also contained in the meta file.

technical Validation
Data inspection. Four inspectors manually performed qualitative data inspection for the datasets. After reconstruction, all images were converted from HDF5 format to MPEG-4 format, at which time co-recorded audio (WAV format) was integrated. All qualitative data inspection was performed manually based on MPEG-4 format videos. Included in the dataset were files that met all of the following criteria: 1. The video (MPEG-4) exists (no data handling failure). 2. The audio recording (WAV) exists (no data handling failure).

The video and audio are synchronized (based on inspection by a human).
After the inspection, 75 participants were included in this dataset. It should be noted that although three participants (sub18, sub74, and sub75) present severe radio-frequency-interference artifacts that were later determined to be from a leak in the MRI radio-frequency cage, we included those three participants in this dataset in the hopes of potentially facilitating development of a digital radio-frequency interference correction method. Files that did not exist (criteria 1) were annotated for each task and participant in the meta file (i.e., metafile_pub-lic_ < timestamp > .json). Figure 2 shows representative examples of the data quality from three participants that are included in this dataset (sub35, sub51, and sub58). Note that all 75 participants exhibited acceptable visualization of all soft tissue articulators. However, two types of artifacts were commonly observed. 1) Blurring artifact due to off-resonance (white arrows, Fig. 2c): This artifact appears as blurring or signal loss predominantly adjacent to air-tissue boundaries that surround soft tissue articulators. It is induced in spiral imaging by rapid changes of local magnetic susceptibility between the air and tissue. This artifact correction is still an active research area, including methods for a simple zeroth order frequency demodulation to advanced model-based 21,22,30 and data-driven approaches 23,55 . At present, we perform zeroth order correction during data acquisition as part of our routine protocol. 2) Ringing artifact due to aliasing (yellow arrows, Fig. 2c): This artifact appears as an arc-like pattern centered at the bottom-right corner outside the FOV. It is caused by a combination of gradient non-linearity and non-ideal readout low-pass filter when a strong signal source appears outside the unaliasing FOV. This artifact does not overlap  www.nature.com/scientificdata www.nature.com/scientificdata/ with the articulator surfaces that are most important in the study of speech production. However, avoiding and/ or correcting this artifact would improve overall image quality and potentially enable the use of a smaller FOV. Figure 3 shows representative examples of the diverse speech stimuli that are included in this dataset. The intensity vs. time profiles visualize the first 20 seconds of four representative stimuli from sub35, shown in Fig. 2a. These examples show the variety of patterns and speed of movement of the soft tissue articulators observed within a speaker as a function of the speech stimuli performed. Figure 4 contains a histogram of the speaking rate in read sentences 56 . The overall mean speaking rate was 149.2 ± 31.2 words per minute. The statistics are calculated from the read "shibboleth, " "rainbow, " "grandfather [1][2], " and "northwind [1][2]" stimuli for all speakers. These selected stimuli are composed of read full sentences. Figure 5 illustrates variability within the same speech stimuli across different speakers. The image time profiles correspond to the period of producing the first sentence "She had your dark suit in greasy wash water all year" in the stimulus "shibboleth" from sub35 and sub41. Although both speakers share the critical articulatory events (see green arrows), the timing and pattern vary depending on the speakers.

Regularization parameter selection for image reconstruction for RT-MRI.
We performed a parameter sweep and qualitative evaluation on a subset of the data to select a regularization parameter for the provided reconstructions. Ten stimuli from ten different participants were randomly selected. The regularization parameter λ was swept in the range between 0.008 C and 1 C in a logarithmic scale. Here, C represents the maximum intensity of the zero-filled reconstruction of the acquired data. Figure 6 shows a representative example of the impact of λ on the reconstructed image quality. A small λ (=0.008 C) exhibits high noise level in the reconstruction (top, Fig. 6). A higher λ (=0.8 C) reduces the noise level but exhibits unrealistic temporal smoothing as shown in the intensity vs. time profiles (yellow arrows, Fig. 6). The optimal parameter 0.08 C was selected by consensus among four experts in the area of MRI image reconstruction and/or speech production RT-MRI. Once the λ was optimized for the subset of the data, reconstruction was performed for all datasets. We have empirically observed that the choice of λ appears robust across all of the datasets.

Usage Notes
Several papers have been published by our group in which methods are directly applied to a subset of this dataset for the reconstruction and artifact correction of the RT-MRI data. These include auto-calibrated off-resonance correction 22 , deblurring using convolutional neural networks 23,57 , aliasing artifact mitigation 58 (the artifact marked by the yellow arrows in Fig. 2c), and super-resolution reconstruction 59 .
Additionally, several tools have been developed at our site for the analysis and modelling of reconstructed real-time MRI data. These include a graphical user interface for efficient visual inspection 16 and implementations of grid-based tracking of air-tissue boundaries [60][61][62] , region segmentation and factor analysis 63-65 , neural network-based edge detection 66,67 ; region of interest (ROI) analysis [68][69][70] and centroid tracking 71 . Some of these tools are also made available; see code repository at https://github.com/usc-mrel/usc_speech_mri.git as well as Ramanarayanan et al. 3 , Hagedorn et al. 4 , and Toutios et al. 24,72 for detailed reviews.
Our group has developed data-driven computational models including an anatomically-guided factor analysis approach of vocal tract contours derived automatically from RT-MRI to provide a compact representation of vocal tract shapes allowing utterances to be decomposed into a linear combination of time functions for the weights of linguistically interpretable vocal-tract deformations 64 . We also developed the Convex Hull www.nature.com/scientificdata www.nature.com/scientificdata/ Convolutive Non-Negative Matrix Factorization approach to learn and decompose temporal patterns in multivariate time-series data such as vocal tract constriction trajectories 73 . We have deployed computational modeling in specific studies including: (i) Development of a framework for determining the test-retest repeatability of quantitative speech biomarkers across experimental conditions and tasks, and sharing test-retest RT-MRI data freely for research use 18 ; (ii) Investigation of synergies for different constriction tasks in VCV sequences with a RT-MRI biomarker computed using a statistical model of vocal tract forward kinematics 65 , supporting the claim that inter-articulator coordination differs depending on the active synergies; (iii) Development of a new vocal tract length estimation model 74 that affords normalization of interspeaker variability and facilitates acoustic comparisons across speakers; (iv) Investigation of speed-accuracy tradeoffs (Fitts' law) across syllable position and subjects, which led to a theoretical account within models of speech motor control 75 ; (v) Design of a modular architecture for articulatory synthesis from a gestural specification comprising relatively simple models for the glottis, aero-acoustics, and the vocal tract 76 .
RT-MRI vocal tract imaging data also have facilitated a range of linguistic studies focused on the composition and internal structure of speech sounds such as for example the production of liquid-rhotic and lateralmulti-gesture segments. The studies have probed the articulation of English liquid consonants in syllable onsets and codas, comparing tongue posture produced in syllable margins adjacent to different vowels 77 . Laterals exhibited greater temporal and spatial independence between their tongue tip and tongue rear gestural components, whereas rhotics were produced with a variety of speaker-specific postures but showed less variation across syllable positions than laterals. Harper et al. 78 examined the multiple component gestures of American English/ɹ/-palatal, pharyngeal, and labial-to understand each gesture's relative contribution to third formant lowering. Data from the USC-TIMIT RT-MRI corpus 16 were analyzed to determine the location, length, and aperture for each constriction gesture of /ɹ/, with formant values extracted at the time of maximal constriction for each gesture. Results suggest that each gesture's contribution to F3 depends on between-and within-speaker variation in the articulation of /ɹ/, as variation in the length and location of the palatal and pharyngeal constrictions is associated with differences in their relative effect size on F3 lowering. In another study, Harper et al. 79 observed RT-MRI data for L2 rhotic production for English learners, finding that the component pharyngeal gesture is particularly difficult for L2 learners whose L1 rhotics lack such a constriction, such as French and Greek. Finally, Monteserín et al. 80 compared the articulation of coda rhotic taps and laterals in Puerto Rican Spanish, which often fail to be discriminated by speakers of other Spanish dialects, and found support for the hypothesis that they are produced with distinct articulatory gestures. Oh et al. 81,82 use RT-MRI to examine glottalic consonants in Hausa. These non-pulmonic consonants such as ejectives and implosives are produced by manipulating pressure in the supralaryngeal vocal tract by means of changes in the vertical larynx position-lowering for implosives and raising for www.nature.com/scientificdata www.nature.com/scientificdata/ ejectives. RT-MRI has for the first time provided quantitative descriptions of these larynx actions and their intergestural timing with their critically coordinated oral constrictions. This study finds that implosives and voiced stops contrast in their internal intergestural timing, not in magnitude of larynx lowering. The observed coordination between vertical larynx movement and the oral closure can be analyzed as in-phase coupling for pulmonic stops  Fig. 4 Histogram of words per minute during scripted speech stimuli including "shibboleth, " "rainbow, " "grandfather [1][2], " and "northwind[1-2]". www.nature.com/scientificdata www.nature.com/scientificdata/ but anti-phase (sequential) coupling for glottalic consonants, implying that in addition to the combination of gestures that comprise a segment, internal phasing relations (or timing) of these gestures must also be encoded in the phonological representation of these complex gestural molecules. In a complementary study, Proctor et al. 77,83 use RT-MRI to illuminate the production of click consonants by a speaker of Khoekhoegowab, revealing that different articulatory actions are used to rarefy the mid-oral cavity in different clicks. Dental clicks are produced with tongue body lowering with a laminal tongue front constriction, while alveolar clicks are characterized by a more apical, rapid tongue tip release with the tongue rear remaining raised in the uvular region.
Finally, the RT-MRI methods have also inspired studies focused on the nature of vocal tract shaping in specific clinical contexts. Lander Portnoy, Goldstein, and Narayanan 84 documented the speech of a patient who has undergone partial glossectomy, at both pre-operative and post-operative time points. They used RT-MRI to observe changes in the variability and principal axes of movement for the tongue. Hagedorn et al. 85 presented an RT-MRI analysis of speech errors in a speaker with acquired apraxia, capturing two types of gestural intrusion errors in a word pair repetition task, as well as intrusion errors in nonrepetitive speech, multiple silent initiation gestures at the onset of speech, and covert (unphonated) articulation of entire monosyllabic words. Lastly, McMicken et al. 86 and Toutios et al. 87 employed RT-MRI to probe the speech production patterns of a speaker with congenital aglossia, a rare syndrome in which an individual is born without a full tongue, who nonetheless has acquired the ability to produce highly intelligible speech. The study found that the speaker, in the absence of a tongue tip, produces plosive consonants that are perceptually similar to /t, d/ by using a bilabial constriction with a significantly increased anteroposterior extent (relative to her /p, b/ productions), accompanied by an increase in closure duration and a widening of the pharynx 87 .

Code availability
This dataset is accompanied by a code repository (https://github.com/usc-mrel/usc_speech_mri.git) that contains examples of software and parameter configurations necessary to load and reconstruct the raw RT-MRI in MRD format. Specifically, the repository contains demonstrations to illustrate and replicate results of Figs. 2-6. Code samples are available in MATLAB and Python programming languages. All software is provided free to use and modify under the MIT license agreement. For a smaller λ (=0.008 C), the reconstruction shows a higher noise level and obscured articulatory event (green arrows), whereas for a higher λ (=0.8 C), the noise level decreases but the temporal smoothing artifact is evident in regions indicated by yellow arrows. λ = 0.08 C yields an acceptable noise level while showing adequate temporal fidelity and therefore was selected for the optimal value for the reconstruction.