Electromagnetic articulography (EMA)1 is a point tracking method, whereby sensors placed on target articulators (including tongue, lips, and jaw) are used to track movement in real time in 3D. As with any method, there are both advantages and disadvantages to EMA (Kochetov, 2020; Earnest & Max, 2003; Maeda et al., 2006; Mennen, Scobbie, de Leeuw, Schaeffler, & Schaeffler, 2010; Stone, 2010; Whalen et al., 2005). We first discuss some advantages of EMA. The data collected within the oral cavity has high spatial accuracy and temporal resolution (see Section 2.4 below), yielding relatively precise information on articulatory gestures. Unlike with some other methods (such as ultrasound tongue imaging), it is possible to measure multiple articulators simultaneously and therefore allows the investigation of inter-articulatory interactions. It is one of the few methods that allows researchers to study movements of articulators directly, as opposed to more indirect acoustic methods. EMA is biologically safe (contrary to some methods used in the past, such as x-ray cineradiography or microbeam) and minimally invasive. Furthermore, the sensors are mostly well-tolerated by adult participants and only moderately interfere with speech production (speakers adapt within 10 minutes; Dromey, Hunter, & Nissen, 2018). Compared to other methods used to track speech articulators, articulographs restrict the participants’ movement less, they are not line-of-sight (such as, e.g., VICON or OptoTrak), and they are not restricted to in-plane visualization (such as, e.g., real-time magnetic resonance imaging or ultrasound tongue imaging).

However, several limitations should be considered when employing EMA for speech-related investigations. For example, the positioning of sensors is limited to the anterior oral tract. It is more problematic to place sensors on the more posterior part of the tongue (e.g., tongue dorsum) than its anterior part, and it is not possible to track velum movements without discomfort to the participants (see exceptions below). Furthermore, depending on the size and location of the articulator of interest, it is not possible to place many sensors on an articulator at the same time due to mutual electrical interference and increased perturbation of articulation. Additionally, sensors still cannot be placed too close to each other without disturbing their measurement accuracy (the Carstens AG500 manual, for example, states that the minimum distance between sensors should be 8 mm), which again limits the number of points that can be tracked on the articulators. Furthermore, because EMA is a fixed point-tracking technique, it does not capture the global movements of articulators, for instance the full midsagittal tongue shape (as obtained using rtMRI).

Additionally, the equipment is expensive and requires a relatively high level of technical knowledge, prior training, and practice to use successfully. Finally, as sensors are firmly affixed to orofacial structures, they constitute a form of articulatory perturbation. While articulation does return to nearly normal after a while (see below), the acoustics are changed when sensors are attached (Meenakshi, Yarra, Yamini, & Ghosh, 2014). Nevertheless, some earlier problems (such as restricted head movement, the need for extensive calibration, and data being restricted to the midsagittal plane only) were present for previous articulographs, but have largely been eliminated with the newer devices (see more details below).

EMA systems have been used for speech-related research since the 1980s (see Figure 1 for an overview of EMA market releases). In the past, the MIT system articulograph (Perkell et al., 1992), the Movetrack system (Branderud, 1985), and the Aurora system (NDI; Kröger et al., 2000) were used as some of the first available commercial articulographs.2 For the past two decades and up until recently, there were two main manufacturers with a continuing production of EMA devices, namely Carstens Medizinelektronik (Bovenden, Germany) and Northern Digital Inc. (Waterloo, Canada). Carstens Medizinelektronik has manufactured several articulography devices over time spanning from the late 1980s until now, including models AG100, AG200, AG500, and the most recent AG501. Northern Digital Inc. (NDI) has manufactured the Wave articulograph, which came to the market in 2009 and was discontinued with the arrival of their latest articulograph, the NDI Vox in early 2020. The NDI Vox has since then likewise been discontinued, as NDI decided to reduce their product portfolio (Northern Digital Inc., 2020). Consequently, at present only Carstens offers a commercial articulograph that has not been discontinued.

As articulographs are costly, it is not uncommon for a lab to use an older system despite a new version being available on the market. Regardless, considerable advancements have been made since the first commercial articulograph. Technological advances have made it possible to collect more comprehensive data, going from 2D EMMA (midsagittal) systems to 3D (or rather 5D) systems collecting three Cartesian coordinates and two angular coordinates (Hoole & Zierdt, 2010). Thus, although early articulographs only measured in one plane (i.e., the midsagittal plane), modern devices track data in three isotropic spatial and two angular dimensions, and sensor orientation is tracked in addition to position. Furthermore, early articulographs required extensive calibration before testing and restricted the participants’ head movement, while modern systems permit free head movement.

Starting in the 1980s, EMA was designed as a way to track points both inside and outside the vocal tract (Schönle et al., 1987). Early studies evaluated the suitability of EMA for tracking speech movements (e.g., Höhne et al., 1987; Hoole & Gfoerer, 1990; Maurer, Gröne, Landis, Hoch, & Schönle, 1993) as well as for clinical use (e.g., Schönle, Müller, & Wenig, 1989; Engelke, Schönle, Kring, & Richter, 1989; Engelke, W., Engelke, D., & Schwetska, 1990). Nowadays, EMA is predominantly employed for the study of speech motor control—in individuals with and without speech disorders—but its uses remain broad. For example, it can be used for the study of orofacial processes in which articulators are actively involved, such as mastication (e.g., Peyron, Mioche, Renon, & Abouelkaram, 1996; Fuentes et al., 2018; Hoke et al., 2019) or swallowing (e.g., Horn, Kühnast, Axmann-Krcmar, & Göz, 2004; Steele & van Lieshout, 2009; Alvarez, Dias, Lezcano, Arias, & Fuentes, 2019; see also Steele, 2015, for a short overview of EMA and other instrumental techniques for the study of swallowing).

The uses of EMA in the study of speech production are likewise varied. Beyond collecting parallel acoustic data, there has been a continued interest in supplementing articulographic data with other speech data, either by collecting data with two devices simultaneously in the same session (if technically possible) or by collecting data from the same participants in separate sessions and coupling the data afterwards. Some of the methods that have been used to collect data in the same session as EMA include: ultrasound tongue imaging (UTI) (e.g., Aron et al., 2016; Benuš & Gafos, 2007), electropalatography (EPG; West, 1999; Simonsen, Moen, & Cowen, 2008; Harper, Lee, Goldstein, & Byrd, 2018), electromyography (EMG; e.g., Rong, Loucks, Kim, & Hasegawa-Johnson, 2012), and motion capture (e.g., Kroos, Bundgaard-Nielsen, & Best, 2012; Krivokapić, Tiede, & Tyrone, 2017). EMA and UTI, especially, are frequently used together, as EMA sensors can be used to provide a fixed reference for ultrasound recordings (e.g., Tiede, Chen, & Whalen, 2019). Methods whose data can be coupled with EMA data after recording additionally include real-time magnetic resonance imaging (rtMRI; e.g., Kim, Lammert, Ghosh, & Narayanan, 2014). Successful attempts have also been made to collect data from two speakers simultaneously using a dual EMA setup (e.g., Geng et al., 2013; Tiede et al., 2010).

Some researchers have made their EMA databases publicly available, sometimes concurrently with other kinematic data collection methods (e.g., rtMRI and UTI data collected from the same participants). Notable articulatory corpora include the USC-TIMIT multimodal speech production database (Narayanan et al., 2014), the MOCHA-TIMIT multi-channel articulatory database (Wrench, 2000), the TORGO database of acoustic and articulatory speech from dysarthric speakers (Rudzicz et al., 2012), the EMA-MAE corpus of Mandarin-Accented English (Ji, Berry, & Johnson, 2014), the mngu0 articulatory corpus (Richmond, Hoole, & King, 2011), the Haskins rate contrast database (Tiede et al., 2017), the MSPKA articulatory corpus of Italian (Canevari, Badino, & Fadiga, 2015), the DKU-JNU-EMA database on Mandarin and Chinese dialects (Cai et al., 2018), the Mandarin-Tibetan speech corpus (Lobsang et al., 2016), and the database of Norwegian speech sounds (Moen, Gram Simonsen, & Lindstad, 2004).

EMA has been used to provide accurate information on movements inside the vocal tract for animating talking heads (e.g., Badin, Tarabalka, Elisei, & Bailly, 2010; Gilbert, Olsen, Leung, & Stevens, 2015), synthesizing speech (e.g., Bocquelet, Hueber, Girin, Savariaux, & Yvert, 2016) or acoustic-to-articulatory inversion (e.g., Girin, Hueber, & Alameda-Pineda, 2017; Sivaraman, Espy-Wilson, & Wieling, 2017), and improving automatic speech recognition (ASR) software (e.g., Demange & Ouni, 2011; Wang, Samal, Green, & Rudzicz, 2012; Mitra et al., 2017). It can additionally be used to provide real time video feedback of articulatory movements and thus has advantages in second language acquisition to help with target pronunciation (Suemitsu, Dang, Ito, & Tiede, 2015) as well as in speech therapy as a biofeedback device (Murdoch, 2011; van Lieshout, 2007). Katz, Carter, and Levitt (2007), for example, used EMA for treatment of buccofacial apraxia, McNeil et al. (2010) used it to study acquired apraxia of speech, and Yunusova et al. (2017) used it to provide feedback to patients with Parkinson’s disease.

Since the advent of EMA devices on the market, their sampling rate and number of channels have increased, and the accuracy has improved. Regarding the recording capabilities of the most recent articulographs, the NDI Wave and NDI Vox have a maximum sampling rate of up to 400 samples/s and can track 16 channels simultaneously (i.e., up to 16 sensors can be used). The AG500 can record 200 samples/s in 12 channels, while the AG501 can record 1250 samples/s of up to 24 channels (Sigona et al., 2018; Savariaux, Badin, Samson, & Gerber, 2017). The speed of current devices is more than enough to capture speech movements from the articulators. For example, Tasko and McClean (2004) indicated that the maximum speed of the tongue body during connected speech was 200 mm/s, and controlled (non-ballistic movements are much slower). A sampling rate of 400 Hz thus has sufficient temporal resolution to track the fastest known articulatory movements.

Several studies have investigated the spatial accuracy of articulographs. Berry (2011) reported that the Wave system showed < 0.5 mm errors for 95% of position samples recorded during human jaw movement for nine out of ten participants. A study on the Carstens AG500 has reported a median error of < 0.5 mm across different types of recordings, including manual movements and various speech tasks, with the error magnitude being dependent on calibration and on the location of the sensors in the electromagnetic field as well as on the proximity between the sensors (Yunusova, Green, & Mefferd, 2009). In addition, the AG500 was found to display some numerical instabilities and anomalies (Stella, M., Stella, A., Grimaldi, & Fivela, 2012) which were not predictable (Kroos, 2012). Finally, a comparison between the Wave and several Carstens systems (namely the AG200, AG500, and AG501) revealed that all four devices showed a local precision of around 1 mm, but a large range of global precision, spanning from 3 mm to 21.8 mm (Savariaux et al., 2017), with the AG501 as the most accurate device with precision of 0.3 mm (RMS; Electromagnetic Articulograph, 2019). Comparisons of the AG500 and AG501 additionally revealed that the AG501 was found to be more accurate, stable, and user friendly (Stella et al., 2013; Sigona et al., 2018) than the AG500. A recent study on the newest NDI articulography—namely, the NDI Vox, which has been discontinued recently—has shown it to be significantly more accurate than the NDI Wave, with an average sensor pair tracking error of 0.1 mm, although a direct side-by-side device comparison would be necessary to establish how the Vox compares with the AG501 (Rebernik, Jacobi, Tiede, & Wieling, in revision).

In general, electromagnetic articulographs are safe to use (Hasegawa-Johnson, 1998). The AG500, AG501, NDI Wave, and NDI Vox articulographs fulfil the safety requirements for electrical equipment as set by the International Electrotechnical Commission and the American Federal Communications Commission (Carstens AG500 Manual, 2006; Carstens AG501 Manual, 2014; Wave User Guide, Northern Digital Inc., 2009, rev. 2016; Vox User Guide, Northern Digital Inc., 2019). Note, however, that little research has been targeted specifically at the electromagnetic frequency ranges of EMA systems (Hoole & Nguyen, 1999; Earnest & Max, 2003). Furthermore, due to the moderate strength magnetic field3 a few exclusion criteria must be considered that impact participant recruitment, predominantly the use of implanted devices that might be prone to electromagnetic interference. These include (as discussed in the Wave User Guide, Northern Digital Inc., 2009, rev. 2016, and Carstens AG500 manual, 2006):

Some studies have tested the potential adverse effects of the EMA magnetic fields on metal objects in the field and, vice versa, the effect of metal objects on the integrity of the collected EMA data. Katz et al. (2003) tested compatibility of the Clarion 1.2 S-Series cochlear implant with the Carstens AG100 articulograph in order to determine whether EMA affects the functioning of the implant and the participants’ speech perception on the one hand, and whether the implant could potentially affect the accuracy of EMA data on the other hand. They determined that the tested cochlear implant was compatible with the AG100, as no adverse effects could be observed.

Joglar, Nguyen, Garst, and Katz (2009) tested potential interference between pacemakers/implantable cardioverter-defibrillators with the Carstens AG100. They determined that devices from Medtronic (type D154VRC), St. Jude (types 5172 and V-193), and Guidant (types 1860, T180, 1852 and 1853) were compatible with the Carstens AG100. Finally, Mücke et al. (2018; see also Hermes, Mücke, Thies, & Barbe, 2019) tested Essential Tremor patients who had undergone thalamic deep brain stimulation (DBS) surgery. Participants were tested using the Carstens AG501 while the implant was active and inactive, with no reported adverse effects. However, as new articulographs and medical devices are introduced, it is necessary to verify their field strength and electromagnetic frequency before doing any testing on participants. Additionally, some researchers advise against including pregnant women in empirical studies using EMA (Hoole & Nguyen, 1999; Stone, 2010) as the effect of the magnetic field is not entirely clear and it is better to err on the side of caution.

Due to the high time demands of the method—including long participant preparation times as well as data processing and analysis steps—EMA studies frequently limit their number of participants. Our literature review (see description below) showed that around 75% of studies published in journals included ten participants or fewer; around 46% included five participants or fewer. This is also in line with Kochetov (2020), who reported the median number of participants in an EMA study to be five. Early studies (e.g., earlier than 2003) have often only included one or two participants, and it was not uncommon for one of the authors to be a participant. With EMA’s increasing popularity, however, there has also been an increase in the number of studies with more participants, with the largest participant samples including around 50 participants (e.g., Schötz, Frid, & Löfqvist, 2013, N = 50; Cheng, Murdoch, Goozée, & Scott, 2007, N = 48; Wieling et al., 2016, N = 48).

In general, most participants tested with EMA are healthy adults (around 80% of the studies). Nevertheless, several studies have tested children from five years of age onwards (e.g., Katz & Bharadwaj, 2001; Cheng et al., 2007; Schötz et al., 2013), giving important insights into the development of individual articulators during the process of early speech acquisition. Articulographs have also frequently been used to study disordered speech in individuals suffering from various conditions that can impact speech production and/or speech motor control, ranging from speech disorders such as stuttering and cluttering (Didirkova & Hirsch, 2019; McClean, Tasko, & Runyan, 2004; Hartinger & Mooshammer, 2008) or apraxia or speech (e.g., Bartle-Meyer, Goozée, & Murdoch, 2009; Nijland, Maassen, Hulstijn, & Peters, 2004); hypokinetic dysarthria (e.g., Kearney et al., 2018; Mefferd & Dietrich, 2019) or Amyotrophic Lateral Sclerosis (e.g., Lee & Bell, 2018; Shellikeri et al., 2016) to congenital conditions such as cleft lip (e.g., van Lieshout, Rutjes, & Spauwen, 2002) or congenital blindness (e.g., Trudeau-Fisette, Tiede, & Ménard, 2017). Using EMA to study disordered speech (more studies can be found in the Appendix) is important to provide insight into the underlying issues of speech motor control that cannot be detected through acoustics only. However, as a method, EMA can also be more fatiguing, and researchers should thus distinguish between what they can and should ask of their participants (Gibbon, 2008; van Lieshout, 2007; see below).

To draw valid conclusions about speech kinematics and speech motor control based on EMA data, it is necessary to ensure between-subjects and between-studies comparability. On the one hand, it is important to correctly place EMA sensors on the speech articulators depending on the specific goals of the study and to optimize sensor adhesion time to ensure cross-trial comparability (after re-attachment, a sensor might not be in the exact same position as before). On the other hand, it is necessary to make the experimental procedure as comfortable as possible for participants while not impeding scientific accuracy.

In the sections below, we lean on our literature review to report some general information on sensor placement, followed by information on certain anatomical considerations that might result in a different sensor attachment strategy, and finally information on the placement and preparation of specific sensor categories (including reference sensors, jaw movement sensors, tongue sensors, and lip sensors).

At this point, we would like to emphasize that most authors follow a certain template when reporting on their EMA study. Such a template is usually of the form:

In the following sections, we discuss the variables that are indicated in this template in bold. Some of the other parts (such as devices and sampling rates) have already been discussed above. Finally, the following sections do not provide information on the EMA data analysis process: The reader is directed to consult Gafos, Kirov, and Shaw (2010) who provided guidelines for using mview, the frequently-used EMA data analysis programme developed by Mark Tiede at Haskins Laboratories (Tiede, 2005); Hoole (2012) who provides a tutorial on his software for processing AG500/AG501 data; and Kolb (2015) who details some other existing software tools and analysis methods. A tutorial on how to analyze EMA data using non-linear regression techniques is provided by Wieling (2018).

Articulographs can be used to study the behaviour of both extraoral (i.e., the lips and the jaw) and intraoral (i.e., the tongue) articulators. The exact choice of sensors depends on several factors, including the studied population (clinical versus healthy, see below; impacts the number of intraoral sensors) and the sounds that are to be investigated (e.g., apical versus lateral; impacts sensor placement). Researcher preference also plays a role: Some prefer to adhere the minimum number of sensors (to decrease the time necessary for participant preparation), while others prefer to adhere more sensors (to collect additional data, using it to answer more research questions). With few exceptions, sensors are almost always placed midsagitally.

The number of intraoral sensors is an important consideration in EMA studies. On the one hand, having more sensors on the tongue allows the tracking of more points and thus yields a better picture of the movement of the tongue. On the other hand, when also including the intraoral jaw movement sensor and reference sensor on the upper and lower incisors, respectively, speakers frequently have five or more wires in their mouth. This may lead to discomfort and affect participants’ speech. More tongue sensors are especially problematic where sensitive populations are concerned. These individuals may be more prone to fatigue (e.g., Friedman et al., 2007, on fatigue in PD patients), more likely to drool (Reddihough & Johnson, 1999), and find it more difficult to stick out their tongue or open their mouth. Furthermore, their speech is more likely to be impeded by a foreign object in their oral cavity. In the case of children, their tongues are smaller, they also salivate more, and need more frequent toilet visits, which necessitates shorter experimental procedures, including shorter preparation times. When testing children and patients, researchers therefore often opt for only two tongue sensors (tongue tip and tongue back) in addition to the intraoral jaw movement sensor and the intraoral reference sensor.

While the exact sensor placement depends on the study, there are some typical sensor placements. These are depicted in Figure 2, which shows movement sensors used to track the movement of articulators (red dots; including the lips, jaw, and tongue) and reference sensors, placed on orofacial structures that do not move during speech production (green dots; including both mastoids, the nasion, and upper incisor). More details on individual sensor categories are provided below.

After all sensors have been placed, a biteplate6 recording can be made with a biteplate object that has several sensors attached to it (see Figure 9 in Section 4.2 for a picture of our lab’s biteplate with three sensors). The object is placed between the participant’s teeth and a recording is made to obtain the relative orientation of the sensors on the biteplate compared to the reference sensors. This information is then used to rotate the acquired sensor movement data (of the sensors attached to the articulators) to a comparable occlusal plane per participant (Westbury, 1994). Finally, palate trace recordings are made, where a sensor is used to trace the palate across the occlusal plane, providing an estimate of the shape of participants’ oral cavity (see Neufeld & van Lieshout, 2014, for a description on how EMA sensors can be used to construct a 3D model of the hard palate).

The time it takes for all sensors to be placed varies. Earnest and Max (2003), for example, state that it can take anywhere between 30 and 60 minutes. This time can be reduced depending on the device, the number of sensors, and their placement. Before starting the experiment, researchers additionally allow some time for the participants to adjust to the sensors. A study by Dromey et al. (2018), who tested sensor habituation, found that after ten minutes, participants reached a level of habituation to the sensors that did not improve even if the habituation stage lasted longer. In general, if researchers include a sensor habituation stage, it is most often 5–10 minutes of informal conversation (e.g., Katz, Mehta, & Wood, 2018; Goozée et al., 2007).

Several brands of adhesive can be used to adhere the sensors. The Carstens website recommends Epiglu (Meyer Haake GmbH), whereas NDI does not give any adhesive recommendations on their website. Other popular adhesives include PeriAcryl®90HV (Glustitch), Isodent cyanoacrylate adhesive (Ellman International), Cyano Veneer Fast (Scheu Dental Technology), Cyanodent (Ellman International), Histoacryl (B. Braun), and Aron Alpha (Toagosei). Note that IsoDent and Cyano Dent adhesives appear to be discontinued7, and Cyano Veneer Fast has not renewed its medical certification, while the intraoral use of Histoacryl may be problematic due to potential cytotoxic effects (Schneider & Otto, 2012). PeriAcryl®90HV has been used most often in recent years.

What these adhesives (except for Histoacryl; Schneider & Otto, 2012) have in common is that they are intended for oral tissue (e.g., for use in dental or oral surgery), are biologically safe, and relatively viscous. Dental cements, including Ketac™, Durelon, and Fuji, have also been used by several labs to attach tongue sensors (e.g., Mooshammer, Hoole, & Geumann, 2006; Tabain, 2003; Steele & van Lieshout, 2004), but are more invasive, as they involve covering the tongue dorsum with a hard substance. Dental cement also causes faster deterioration of sensors and leads to participant discomfort. However, it does have the benefit of making sensors adhere to the tongue for a longer period of time (e.g., Ball, Gracco, & Stone, 2001, state that the sensors remain firmly attached to the tongue surface for over 90 minutes).

Before discussing frequent sensor placements, it is also necessary to mention some more unusual sensor placements. In the past, sensors have been adhered to the velum using different means, from glue to atraumatic sutures (e.g., Engelke, Hoch, Bruns, & Striebeck, 1996, number of participants N = 1; Okadome & Honda, 2001, N = 3; Jaeger & Hoole, 2011, N = 4). Other orofacial structures to which sensors have been adhered include the uvula (e.g., Hoenig & Schoener, 1992, N = 30), thyroid cartilage/skin above the larynx (e.g., Alvarez et al., 2019, N = 14; Shosted, Carignan, & Rong, 2011, N = 4; Bückins, Greisbach, & Hermes, 2018, N = 4), and sublaminally on the underside of the tongue (e.g., Rochon & Pompino-Marschall, 1999, N = 4).

In the procedure used in our lab, all sensors are prepared at least half a day before the experiment. In this preparation stage, we distinguish between three types of sensors: (1) the extraoral sensors (identified with MR, ML, N, UL, and LL, below) plus the sensors attached to the tongue (TM and TT), except for the most posterior tongue sensor, (2) the most posterior tongue sensor (TB), and (3) the sensors attached close to the incisors on the upper and lower gums (UI and LI). We check the sensors for any visible defects (e.g., broken wire) before using them.

The first group of sensors is prepared by dipping each of them in mask-making latex (RD 407 Mask Making Latex, Monster Makers). The TB sensor is prepared similarly but having an additional latex flap cover (see Section 5.1), which increases the surface of the sensor and may be beneficial for the adhesion duration (see Section 5.5). Finally, the UI and LI sensor are prepared using a Stomahesive wafer (ConvaTec PLC). A small rectangular piece of Stomahesive is cut measuring about 10 mm × 6 mm. The sensor is placed on top of this piece and a drop of latex is applied to it in order to make it adhere (Figure 4, left and right). The early preparation phase is necessary, as the latex takes several hours to completely dry. However, the sensors should not be prepared too early (e.g., a week in advance), as the latex becomes less flexible with time and more difficult to remove. In case of re-use, we disinfect sensors first using SPORECLEAR Medical Device Disinfectant (Hu-Friedy Mfg. Co., LLC) and then wipe them with an alcohol wipe before storing them.

After checking that participants are not pregnant, do not have a pacemaker, and do not have a latex allergy, our data collection procedure is as follows. All sensors are screwed into the miniature terminal blocks of the NDI Wave (or, in the case of the NDI Vox, plugged into the sensor harness assembly), wiped with an alcohol wipe, and placed on a sterilized tray a short time before the participant’s arrival. We perform a sensor validation check by verifying that each sensor that is screwed in also functions as it should. Once participants arrive, we first ask them to take a disposable toothbrush and scrub their tongue (especially along the midline). They do this in front of a mirror, so that they are aware of how far back they are reaching and do not trigger their gag reflex. By scrubbing their tongue, they remove the coating that covers the tongue (the amount of coating differs per participant10). We subsequently ask the participant to remove jewellery, glasses, and hearing aids, when applicable, as they make sensor placement more difficult and potentially could interfere with the signal (as the presence of metal inside the magnetic field has a negative effect on the precision of the recovered sensor positions). The glasses and hearing aids are returned to the participant once the sensor placement is complete if their use is necessary for successful participation in the experiment.

We additionally ask participants whether they are wearing dentures, as these may move slightly during speaking, which could result in some wire pull for sensors placed on the gingival tissue. Since dentures cannot be removed without impeding articulation, we note their presence but otherwise do not ask the participant to remove them. Additionally, if possible, participants should shave before the experiment and avoid wearing makeup as this makes sensor placement more difficult.

Subsequently, the participant is asked to sit down next to the EMA field generator (we were using the NDI Wave system, but have very recently moved to using the NDI Vox system). We first place four prepared reference sensors:11

All sensors (reference and others) are first held in reverse action tweezers (Hobbycraft), as they make the application of sensors to the participant easier. The first three reference sensors are applied after the researcher has sterilized their hands using Sterilium® (Medline). Before placing any intraoral sensors, the researcher puts on (latex) dental gloves and a dental mask.12

The mastoid sensors (ML and MR) are placed behind the participant’s ears on the skin covering the mastoid part of the temporal bone, where there is minimal skin movement (Figure 5).

The nasion sensor (N; Figure 6) is placed on the part where there is least skin creasing. If the participant is wearing glasses, the sensor is placed right above or below their glasses, depending on how big the frame is. The first three sensors are secured with a drop of glue. We use PeriAcryl®90 HV adhesive (GluStitch Inc), which is kept in the fridge (at ~2°C) until the participant’s arrival. At that moment, two to three drops of adhesive are added to a small plastic mixing well (Maxill Inc.) after which the adhesive is returned to the fridge. A small disposable plastic pipette is used to transfer the adhesive from the mixing well to the sensor.

The sensor wires are adhered to the participant using Leukopor or Leukosilk tape (BSN medical GmbH). A piece of tape is additionally placed over the ML and MR sensors to secure them (see tape in Figure 5). We add a piece of tape to the N sensor but place it slightly higher on the forehead (see tape in Figure 6), as it otherwise disturbs the participant’s visual field.

The final reference sensor (UI), on top of the piece of Stomahesive, is attached to the gingiva above the left upper incisor. No glue is added to the Stomahesive, as it adheres to tissue by itself. We avoid placing any incisor sensors to the midsagittal line, directly above the central incisors, due to the labial frenulum, which connects the upper lip to the gingival tissue and is quite sensitive. The UI sensor placement relative to the labial frenulum can be seen in Figure 7.

After the reference sensors have been placed, the palate trace and biteplate recordings follow. These are crucial (particularly the biteplate recording) to ensure the subsequent quality of the collected data. For the palate trace, we adhere one spare sensor to the end of the participant’s dominant thumb using Leukopor tape (so that the sensor wires are leading down the thumb and pointing towards the wrist) and instruct them to trace the thumb from the back of the hard palate to their front teeth. The purpose of this procedure as well as the tracing method are explained by means of a mouth puppet (Super Duper® Publications; Figure 8), which, due to its cartoonish look, is also useful in decreasing participants’ potential anxiety. The palate trace is performed twice.

For the biteplate recording, we created a (reusable) fixed triangular protractor with three sensors glued to it (Figures 9 and 10). The same protractor is used for all participants; it is wiped with an alcohol wipe before every use and disinfected with SPORECLEAR Medical Device Disinfectant (Hu-Friedy Mfg. Co., LLC) after every use. The protractor is pushed as far back as comfortable into the corners of the participant’s mouth. The participant is then asked to hold the protractor firmly between their teeth and sit still for a few seconds while the biteplate recording is made. The protractor must be in contact with the molars in order to obtain a true occlusal reference. We check the biteplate recording directly by comparing the Euclidean distances between all the reference sensors and the three sensors on the biteplate, using MATLAB (MathWorks Inc.). If these distances remain relatively constant over time, this indicates that the position of the reference sensors and the biteplate sensors is correctly tracked.

After the palate trace and biteplate recordings, we proceed with attaching sensors to the articulators that we wish to capture. Most frequently, these sensors are the following (listed in the order of placement):

To determine where to place the tongue back sensor, we use a colour transfer applicator stick (Dr. Thompson’s, GUNZdental). We ask the participant to drag the stick midsagitally across the midline of their hard palate (as they had done before with the palate trace sensor) and then pronounce the velar /k/, followed by directly sticking out their tongue.13 They are asked not to swallow while their tongue is being marked. The colour from the applicator is transferred from the palate to the part of the tongue where the back-most (velar) sound is made. We use the same stick to draw a coronal line through this spot. Additionally, we use measuring tape to measure 1 cm from the tongue tip (when the tongue is stretched) and drag a coronal line through that point as well. The coronal line enables us to always re-adhere the sensor to the approximately same position if it starts getting loose, as the point might become smudged through speaking and swallowing, but the line will remain clearly visible. Figure 11 below shows the coronal lines on the tongue left by the colour transfer applicator stick, with the median sulcus still clearly visible.

The participant can now swallow as the coronal lines will remain clear, even when they come in contact with saliva. The participants are asked to stick out their tongue as far as comfortable. We place barber tape (Comair GmbH, folded three times to contain at least eight layers) on the back line marking on participant’s tongue, dab the tape on the tongue for about 5–10 seconds, and finally drag the tape across the tongue. This procedure dries the tongue dorsum and is crucial in ensuring that sensors do not fall off easily. We hold each sensor in the tweezers and add a drop of adhesive using a small plastic disposable pipette before placing the sensor on the tongue.

The TB sensor is placed on the crossing between the marked posterior line and the median sulcus, so that the wire of the sensor is pointing downward and towards the lip corner. A disposable wooden tongue depressor (Tegler) is used to press the sensor to the tongue for 10–20 seconds. The wire is then secured to the cheek using Leukopor tape. It is essential that the wires have enough slack, as large speech gestures may otherwise lead to wire tension, which is uncomfortable for the participant and may cause the sensor to come loose. The process is repeated for the TT sensor, which is placed on the crossing between the marked anterior line and the median sulcus. Note that the TT sensor is positioned in such a way that the wire is pointed towards the side of the tongue, as a wire running over the tongue tip feels uncomfortable for the participant and leads to lisping (Hoole & Nguyen, 1999).

The tongue mid sensor is placed halfway between the marked lines for the TT and TB sensors on the median sulcus by eyeballing. In line with previous methodological considerations (see Section 3.5.1), we generally do not use the TM sensor when testing clinical populations or children. If we are using lateral sensors, we place these to the right and left side of the TM sensor, 0.5–1 cm from the edge of the tongue (depending on how wide or narrow the participant’s tongue is). We only place more than three sensors if that is required for the purposes of the study. The final intraoral sensor (LI) tracks the jaw movement. This sensor, prepared with Stomahesive, is attached to the gingiva below the right lower incisor. No additional glue is needed, as Stomahesive adheres to tissue by itself.

Finally, two lip sensors (UL and LL) are attached at the vermillion border of the upper and lower lip using a drop of dental adhesive. Depending on the amount of facial hair surrounding the upper and lower lip, the removal of lip sensors can lead to mild discomfort.

The present experiment tested how different preparation methods for EMA sensors affect adherence to the tongue. As discussed in Section 3.5.2, several methods for sensor preparation exist. We specifically focus on the tongue sensors, as these usually are most likely to come off relatively quickly. The aim of this experiment was therefore to determine which type of sensor preparation (see below) is most beneficial for adhesion, also depending on the position on the tongue. In addition, we evaluated (qualitatively) whether the participant’s tongue anatomy influences adhesiveness.

We tested three types of sensor preparations: out-of-the-box (‘bare’) sensors, latex-coated sensors, and sensors with a latex flap. Out-of-the-box sensors (Figure 12, left) are the sensors as provided by NDI for the Wave device (approximate surface: 30 mm2), latex-coated sensors (Figure 12, center) are dipped in latex (with only a slightly larger surface than the out of the box sensors, but with rounder edges), and sensors with a latex flap (Figure 12, right) are covered in the same latex, but now a brush is used to apply the latex while the sensor head is lying on a flat surface (approximate surface: 70 mm2).

To test these three types of sensor preparations, we tested 10 female adult participants in three separate sessions. All 10 participants were between 20 and 30 years of age. The study was approved by the Faculty of Arts Research Ethics Review Committee of the University of Groningen (approval number 71276154).

For each of the three sessions, we used one type of sensor and followed the same application procedure for each type (as described in Section 4 above). The sessions took place on three different days, thus avoiding the risk of glue residue and tongue fatigue, both of which would have influenced the resulting adhesion times. During the first session, we adhered out-of-the-box sensors, during the second session the latex-coated sensors, and during the final session the sensors with the latex flap.

During every session, we placed five sensors on the tongue, as this is the maximum number of tongue sensors used by researchers (see Section 3.5.1 and the Appendix). The sensors in question were placed on the tongue tip (TT; 1 cm from the tip), tongue back (TB; place of /k/ constriction), tongue middle (TM; between TT and TB), and tongue lateral right and left (TLR and TLL, placed to the left and right of the TM sensor, respectively). While few studies investigate lateral sounds (see above), we wished to assess whether different types of sensor preparations are also suitable for studying lateral movement of the tongue, as those parts move differently, and the sensors are more prone to interference from the participants’ molars. Figure 13 below displays sensor placement examples for latex-coated sensors.

The sensor placement process took approximately ten minutes. After we placed the five sensors on the tongue, we started displaying the stimuli to the participants using Microsoft PowerPoint on a computer monitor in front of them. The articulograph was not turned on for this experiment, as we were not collecting kinematic data and merely wished to determine how long it took for each sensor to fall off.

The experimental procedure consisted of the following tasks and stimuli. First, the participants read the short text Please call Stella from the Speech Accent Archive (Weinberger, 2015). This allowed them to get used to speaking with sensors in their mouth (i.e., the sensor habituation stage) and took approximately one minute. We did not include a longer sensor habituation stage as our goal was not to record the participants’ natural speech. Following the text was a wordlist. It contained 300 words of varying lengths and from various thematic fields (e.g., vegetables, fruit, school, vocations). Each word appeared on the screen for four seconds, during which the participant read it out loud. This procedure lasted for 20 minutes. Finally, at the end of the first wordlist, the participants performed a rapid syllable repetition task, namely the diadochokinesis (DDK) task, at a comfortable but fast speaking pace as defined by the participants themselves. The DDK task involved the repetition of syllables /pa/, /ta/, /ka/, and /pataka/, and was included because fast repetitive movements may potentially cause the sensors to fall off faster. The experiment was considered complete once all the sensors had been detached or the three tasks were repeated twice, which took about 45 minutes. When a sensor fell off (the participants were instructed to inform us when they felt a sensor get loose), we removed it and noted the time it fell off. We did not re-attach any sensors.

The experimental procedure, including participant preparation (five minutes) and sensor placement (ten minutes), took 60 minutes at most. At that point, we stopped the experiment and removed the remaining sensors. The maximum time a sensor was adhered to a person was therefore 45 minutes. The experimental procedure is schematically presented in Figure 14.

For all participants, we measured the relative tongue length, tongue width, and maximal mouth opening. All three measurements were taken with the participant’s tongue comfortably extended using a ruler. First, we measured the relative tongue length, defining it as the distance between the anatomical tongue tip and the place we had marked as the place of /k/ constriction. Second, we measured the tongue width, defined as the widest part of the tongue, parallel to the molars. Finally, we asked the participants to open their mouth as wide as they comfortably could and measured the vertical distance between the surface of the tongue and the edge of their upper central incisors. We defined this as ‘mouth opening,’ which in effect represents the maximum intraoral space that the researcher can work with during the sensor placement procedure.

Due to the lack of suitable equipment, we were not able to measure the participants’ salivary flow rate or take any other anatomical measurements.

To assess the potential effect of sensor preparation method and sensor position on sensor adhesiveness, we used linear mixed effects regression modelling with participant as a random-effect factor and the optimal random-effects structure (i.e., assessing the inclusion of random intercepts and slopes) determined via model comparison. Specifically, we evaluated whether sensor preparation type (OUT-OF-THE-BOX, LATEX-COATED, FLAP) and sensor position (TT, TM, TB, TLL, TLR) affected sensor adhesiveness. As our initial analysis appeared to show a clear distinction between the TB sensor (adhering for a much shorter duration than the other sensors, whose adhesion did not differ significantly from each other; see Figure 15), we created a new fixed effect predictor distinguishing the TB sensor from the other sensors.

The best model for our data, determined via model comparison, only warranted the inclusion of the distinction between the TB sensor and the other sensors, in addition to the by-subject random intercept and a by-subject random slope for the contrast between the TB and the other sensors. Specifically, this model showed that the TB sensor adhered approximately 14 minutes less than the other sensors (β = –14.0, t = –5.0, p < 0.001). Sensor preparation type (see Figure 16) did not reach significance in the best model, nor did any of the other anatomical predictors. Of course, this may be partly due to our limited sample size (N = 10).

When explicitly focusing on the interaction between sensor preparation type and the sensor (TB versus the other sensors), the flap sensor appeared to be detrimental for the adhesion time for non-TB sensors, reducing the estimated adhesion time with about five minutes compared to the bare sensor and about three minutes compared to the sensor coated in latex. However, for the TB sensors the opposite pattern was found. The adhesion time of the sensor with the flap was estimated to be about five minutes higher compared to the sensor coated in latex and more than nine minutes higher compared to bare sensor. Figure 17 shows this interaction; Table 3 shows speaker-specific differences in sensor adhesion times.

In general, our sensor adhesion experiment demonstrated no clear general advantage of any particular sensor preparation type. With five sensors on the participant’s tongue, it was difficult to make all of them adhere for the duration of 45 minutes (the only exception being two participants). The adhesiveness of the TB sensor, which was significantly lower than that of other sensors, did improve when the sensor was prepared with a latex flap. When attaching intraoral sensors, it is crucial to preserve a sterile environment. As sensors coated in latex (both with and without a latex flap) are more hygienic, easier to clean, and likely deteriorate slower, we recommend coating the sensor in latex when possible. Based on our results we further recommend adding a latex flap for the tongue back sensor.

Additionally, we would like to mention some qualitative observations. Placing a total of five sensors on the tongue is not ideal, particularly when keeping in mind that in a regular experiment, two additional intraoral sensors would need to be included as well. This difficulty was especially pronounced with latex flap sensors, as the required tongue surface to attach the sensors to was largest. In the case of using sensors with a latex flap, participants appeared to take longer to get habituated. However, their articulation seemed to return to normal within the first ten minutes of the experiment (although note that we did not quantify this) and should therefore not be problematic for a regular experimental setup. A practical advantage of the sensors with a latex flap was that once part of the flap detaches from the tongue, this is quickly noticed by the participant and can be easily resolved by adding some glue underneath the flap.

There were several limitations to this study. First, we adhered five sensors to the tongue, which is a larger number than usual. While this was done on purpose, as we wished to assess not only adhesion of sensors placed midsagittally but also those placed laterally, it also might not reflect adequately how long the sensors would adhere in a normal experimental scenario with only two or three tongue sensors. Second, we did not readhere the sensors once they fell off. In a real experimental scenario, one would reglue the sensor to the same position where it fell off. In our experience, it is easiest to reglue a sensor with a flap as the adhesive surface is largest.

Another experiment would need to be conducted to assess how the different sensor types compare when focusing on ease of reattachment. Finally, sensor placement and its effectiveness are strongly impacted by individual factors. While we included certain tongue anatomical measures (none of which turned out to be significant predictors in our best model), others that were not measured and differ between participants—such as salivary flow rate (Whelton, 2012) and tongue surface (Kullaa-Mikkonen et al., 1982)—likely play an important role as well.14