Twelve naïve young adults (5 females, 20–40 y.o.) participated in the experiment. They were all French native speakers and had no known speech or hearing impairments. They also had no history of profound injury that could impair somatosensation in the orofacial region. Participants signed the consent form approved by the local ethic committee, CERGA (Comité d’éthique pour la recherche, Grenoble-Alpes) (CERGA-AvisConsultatif-2021-18). One participant was excluded for not following the instructions. Eleven participants were included in the current analysis.

Electromagnetic Articulography (EMA, Wave, Northern Digital Inc.) was used to record articulatory movements in synchrony with the recording of the produced sounds. For the production of a front vowel such as /e/, the articulatory movement is primarily characterized by the tongue position in the mid-sagittal plane. Hence, EMA sensors were attached to the tongue in the mid-sagittal plane (Figure 1A): tongue tip (TT), tongue blade (TB), and tongue dorsum (TD). We planned to set at 1 cm the distance between two neighbor sensors on the tongue. However, this target distance was slightly adjusted in each participant depending on the size of the tongue. Finally, the resultant distances were 1.1 ± 0.68 cm between TT and TB, and 1.1 ± 0.85 cm between TB and TD. Additional sensors were attached to the upper and lower lips (UL and LL), and to the jaw (J), to record potential movements of articulators other than the tongue, which might affect vowel acoustics after the application of tongue stretch. For head movements’ correction in the off-line analysis, four reference sensors were attached to the upper incisors in the mid-sagittal plane, and the nasion and the left and right mastoids. The participant’s head was held in place with a head holder. After each recording, the palate contour in the mid-sagittal plane was recorded by tracing the surface of the palate using an EMA sensor attached to the experimenter’s finger. The sensor data were recorded at a 200 Hz sampling rate, and the speech sound produced was recorded synchronously at a 22.05-kHz sampling rate.

The speech task consisted of the sustained production of the whispered vowel /e/ for about 3.5 s. Vowel /e/ was selected on the basis of the results of our previous study (Ito et al., 2020), in which systematic and large compensatory responses to tongue stretch were observed for this vowel. Since our previous study also showed no reliable difference between voiced and whispered conditions, we selected a whispered production in order to increase the likelihood that the masking noise actually masks the auditory feedback. For this masking, we used a pink noise, which was presented through earphones (Natus Tip 300) at 80 dB SPL. In every trial, the speech task started with the mouth closed and ended by closing the mouth.

For the tongue stretch perturbation, we used the same experimental setup as in Ito et al. (2020) (Figure 1A). A small robotic device (Phantom Premium 1.0, Geomagic) was placed in front of the participant and connected to the tongue surface with a thin thread. The thread has two small anchors, which were attached to the tongue surface on both lateral sides of the tongue blade sensor (TB). The distance between each anchor and the TB sensor was set to be 1 cm on either side of the sensor. This target distance was adjusted in each participant depending on the size of the tongue. The tongue stretch was applied by pulling the tongue forward with a 1 N force for 1 s. The force was applied as a step function, with rise and fall phases of 5 ms to avoid mechanical noise in the robot.

The time sequence of each trial is represented in Figure 1B. Each trial was triggered manually by the experimenter after visually checking that the participants were ready with their mouths closed and with their throats cleared if necessary. The participants carried out the speech task in response to a visual cue (green circle) presented on a monitor. Under the auditory masking condition, the noise was presented from the end of the previous trial to accustom the participants to the noise before the speech task and then lasted to the end of the trial at hand. The onset of the tongue stretch perturbation occurred between 1 s and 1.5 s after the presentation of the visual cue that launched the speech task.

The recording was divided into three sessions. Each session included 5 voiced trials and 50 whispered trials. The participants were asked to speak the vowel /e/ aloud in the first 5 trials and to whisper it in the following 50 trials. The 5 voiced trials were used to make sure that the participants actually produced vowel /e/ (and not vowel /ɛ/ or /œ/). Hence, we did not apply any perturbation during these voiced productions. In the 50 whispered trials, both auditory conditions were tested (25 trials each) in randomized order. The tongue stretch perturbation was applied in a fifth of the pseudo-randomly selected trials, with the constraint that it never occurred in two consecutive trials and that it was applied in the same number of trials in both auditory conditions (5 trials each in one session). In total, 165 trials were carried out (150 repetitions of the whispered speech task and 15 trials with voicing). 15 perturbed trials were recorded in each auditory condition. Despite these precautions, one participant did not correctly sustain the vowel /e/ during the main task and was excluded from the analysis.

Before the main recording, we also carried out a practice session to ensure that the pronunciation of vowel /e/ was correct and not influenced by a regional accent. The participants practiced to whisper the vowel with and without the auditory masking. We also asked the participants to make sure that the level of masking noise actually did mask their whispered sounds properly.

We only analyzed the perturbed trials in the articulatory and acoustic domains. For each trial, time zero was aligned with the onset of the perturbation, and the analysis was applied to the time interval from 1 s before the perturbation onset to 1.5 s after the perturbation onset.

The articulatory movement data were first preprocessed by correcting the movement of the head with the reference sensors. Considering the inter-individual variability in articulatory positioning for vowel /e/ production, we evaluated the relative changes from baseline. Baseline was defined as the average position computed over the 50 ms preceding the perturbation onset, for each sensor and each participant. Since the recorded tongue movements include the influence of jaw movement, we subtracted the position of the jaw sensor (J) from the recorded positions of the tongue sensors (TT, TB and TD). Finally, an average movement signal was computed for all the perturbed trials and all the participants for each sensor and for each auditory condition.

For sound data, the first three formants (F1, F2 and F3) were extracted over the same time interval as for the articulatory data using LPC analysis (Rabiner and Schafer, 1978). In this extraction, the acoustic signals were first under sampled at 10 kHz in order to focus on the frequency range ([0, 5 kHz]) of the first four formants in adult speakers. Sliding time Hanning windows of 25 ms with a shift of 2 ms were used, and an LPC analysis of order 12 was carried out for each window. Four possible formant frequencies were thus extracted at a sample rate of 500 Hz, and the time variations of the first three formants were then computed on the basis of these four frequencies, using basic smoothness and continuity principles. As with the movement data, time zero was aligned with the perturbation onset; the variation of each formant over time was computed for each participant relative to baseline, which was computed as the average value over the 50 ms preceding the perturbation. Finally, an average variation over time was computed for all the perturbed trials and all the participants for each formant and for each auditory condition.

In line with our previous findings (Ito et al., 2020), we expected the response induced by the tongue stretch perturbation to consist of two phases according to the response latency. We interpreted the early phase as the result of the passive mechanical characteristics of tongue tissues, and the later phase as the response induced by neural sensory feedback. In the current study, we focused only on the later phase of response and examined whether the magnitude and timing of the response differ across the two auditory conditions.

Using the displacement, velocity and acceleration along the horizontal direction, we determined relevant time points of interest: the onset of displacement in response to tongue stretch; the time of the maximum displacement in response to tongue stretch; the onset and offset of the compensatory response induced by neural sensory feedback (See Results for the details). The horizontal displacements did not significantly differ between the two auditory conditions as shown in Figure 2B. So the data were averaged over the two auditory conditions and this grand-average was used to determine these time points. All the detected time points are represented by vertical dashed lines in Figure 2.

We measured the amplitudes in the sagittal plane of the initial displacement (from the onset of displacement to the time of maximum displacement, a period called “Ini” henceforth, Figure 2) and of displacement during the compensatory response (a period called “Comp” henceforth, Figure 2). For the sound data, we compared the formant frequencies F1, F2 and F3 at two time points: the time of maximal displacement in response to tongue stretch, and the offset of the compensatory response. For each measure, in the articulatory and acoustic domains, a one-way repeated measures ANOVA was applied to compare between two auditory conditions.

In line with our previous study (Ito et al., 2020), we expected the compensatory movement measured by the TB sensor in the mid-sagittal plane to take the shortest path to the original tongue contour, rather than returning to the exact original position before tongue stretch perturbation. We verified this behavior graphically by estimating the original tongue contour in the sagittal plane as the concatenation of two segments going from TT sensor to TB sensor and from TB sensor to TD sensor. To do this, we averaged the raw sensors’ positions over the 50 ms preceding the stretch onset.

We also assessed whether auditory masking affected the baseline articulatory posture for the production of the vowel /e/, bearing in mind that auditory masking might modify articulatory movement and posture – a phenomenon known as the Lombard effect (Luo et al., 2018). The typical behavior of the Lombard effect is that the energy of the produced sound increases, which is particularly marked in the energy of vowels, mostly without consciousness. We assessed possible consequences of this effect in terms of articulatory posture. We compared the baseline articulatory posture of the tongue including TT, TB and TD between two auditory conditions using repeated measures ANOVA.

We first verified whether the auditory masking affected the baseline articulatory posture for the production of the task utterance in our experiment. We compared the positions of the three tongue sensors (TT, TB and TD) in both auditory conditions. Figure 3 shows a representative example obtained for one participant, which is the baseline articulatory positions in the two auditory conditions. Repeated-measures ANOVA showed no reliable difference between the two auditory conditions (F(1,50) = 0.025, p > 0.8), and no significant interaction effect between sensors and auditory conditions (F(2,50) = 0.016 p > 0.9). This indicates that auditory masking did not affect the basic achievement of the articulatory position for vowel /e/.

Figure 2A shows the horizontal displacement of the TB sensor (top panel) and the corresponding F1, F2 and F3 changes (bottom three panels) observed in response to tongue stretch over time. As in our previous study (Ito et al., 2020), the compensatory response had consequences in the articulatory and acoustic domains alike.

The tongue stretch perturbation first induced a fast forward displacement of the tongue, as characterized by the TB sensor, up to a maximum (first period, called “Ini,” in Figure 2). In a second period of the response, the amplitude of displacement decreased as the result of the combination of passive and compensatory effects. As in our previous study (Ito et al., 2020), we identified three time points of interest to characterize these two periods in the response. The time of maximum displacement characterizes the end of the first period, which is purely due to passive effects. For the second period we relied on the velocity and acceleration profiles, and identified a compensatory response that we consider to be induced by neural sensory feedback. Figure 2B shows a magnified view of the displacement, velocity and acceleration of the TB sensor. We observed an increase in the velocity, which we characterized in time by the peak of acceleration, which corresponds to an inflection point in the displacement. This second time point marks the onset of the compensatory response due to neural sensory feedback. The offset of this compensatory response is characterized by the subsequent velocity zero-crossing, which is the third time point of interest. These points are indicated in both panels of Figure 2 by three vertical dashed lines points of interest. The compensatory response due to neural sensory feedback (“Comp”) is the focus of our analysis.

When the articulatory responses in the two auditory conditions were compared, we observed that the averaged values are mostly similar across participants in all variables. This can be seen in the top panel of Figure 2A, where the shaded areas represent standard-errors across participants for both auditory conditions. The clear overlap of the shaded areas between auditory conditions suggests no significant difference specifically in the Ini and Comp periods. This is quantitatively assessed below with repeated measures ANOVA.

The averaged displacement of TB sensor in the mid-sagittal plane for each auditory condition—from the onset of the perturbation to the offset of the compensatory response induced by neural sensory feedback is presented in Figure 4A. As observed in our previous study (Ito et al., 2020), the tongue was first displaced horizontally in the forward direction due to the horizontal force applied to the tongue, and then the compensatory response occurred. Importantly, as in our previous study the compensatory movement did not follow the path back to the position before the tongue stretch perturbation. Instead it went in the downward direction so as to take the shortest path to a posture that preserved the original tongue contour (as estimated by the dotted lines in Figure 4A) in the alveo-palatal region, in which the constriction of the vocal tract occurs during the production of vowel /e/. In line with our observations in the horizontal direction, the trajectories in the mid-sagittal plane are similar for both auditory conditions. Figure 4B shows the amplitude of the movement response in the mid-sagittal plane for both periods Ini and Comp. The repeated measures ANOVA did not reveal any significant difference between the two auditory conditions in the amplitudes of both periods of the response (F(1,10) = 0.518 p > 0.4 for Ini and F(1,10) = 0.191 p > 0.6 for Comp). The results indicate that the auditory condition did not significantly affect the amplitudes of the initial changes and of the compensatory response.

The articulatory changes observed in response to the tongue stretch perturbation resulted in acoustical changes, as revealed in the variations of F1, F2 and F3 values. The rapid deflection of the tongue observed in the first period of the response induced a rapid decrease of F1 and F2, and a rapid increase of F3. The main effect associated with the decrease of the tongue deflection in the second period of the response is observed on F3 which clearly decreased and tended to return to its value before the perturbation onset. In F1, we essentially observe, for the same period of time, a stabilization with a slight non-significant trend to return to its original value, while F2 continued to decrease, but at a lower rate than in the first part of the response.

The averaged amplitudes of the normalized formant variations during the Ini and Comp periods together with the standard-errors are presented in Figure 5. At first glance, the figure confirms that significant formant changes were induced in the first period of the response to the perturbation (Ini), that only F3 shows significant compensation effects in both auditory conditions during the compensatory response due to neural sensory feedback (Comp), and that F1 does not show a clear trend toward a compensation for the effect of the perturbation. These observations are quantitatively confirmed by the statistical analysis: two-way repeated measures ANOVA show that only F3 underwent a significant change between the onset and offset of the compensatory response Comp [F(1,30) = 11.021 p < 0.003]. There was no interaction effect with the auditory condition for either formants [F1: F(1,30) = 0.358 p > 0.5, F2: F(1,30) = 0.088 p > 0.7, and F3: F(1,30) = 0.001 p > 0.9]. These changes were consistent in both auditory conditions. A one-way repeated measure ANOVA of the amplitude of formant variations during the Ini period showed no significant effect between the auditory conditions for each of the three formants [F1: F(1,10) = 0.276, p > 0.6 F2: F(1,10) = 0.021, p > 0.8, F3: F(1,10) = 0.006, p > 0.9]. Similarly, the two-way repeated measures ANOVA on the onset and offset times of the Comp period show no significant difference between the two auditory conditions in all three formants [F1: F(1,30) = 0.196, p > 0.6, F2: F(1,30) = 0.312, p > 0.5 and F3: F(1,30) = 1.395, p > 0.2]. The absence of difference between auditory conditions indicates that auditory masking did not affect the compensation in response to the tongue perturbation.

Note that a large increase of F1 was only observed after the Comp period in normal auditory conditions. Although this may possibly due to the compensation by auditory feedback, we did not pursue this further since this was beyond the current target hypothesis.