The discrimination task was composed of two lists containing 40 pairs of disyllabic-nonwords, selected from the nonword corpus published by Nielsen and Dau (2019). From the original material, C−/a/-C−/a/ nonwords starting with one of the 14 main initial phonemes for the Danish language (/p t k b d g m n l f v s r h /) were selected. The recordings selected had 100% speech intelligibility score reported in the original study by Nielsen and Dau (2019), for both first and second consonants. The present study focuses only on the first phoneme contrast.

For the task, nonwords were combined in pairs, to account for all the minimal-pair combinations in Danish for which just one distinctive production feature is present (place, voice/aspiration, or manner, as in ‘bafi – ‘pafi’). To facilitate phonological ─ instead of acoustical ─ comparison of the word pairs, recordings of each nonword from three different speakers were selected from the original material and each pair was presented using audio from two different speakers, selected randomly among the three possible recordings. Post-hoc analysis revealed no effect of the speaker-pair used on participants’ performance.

All the audio recordings were normalized by root mean square (RMS) and silence was added in the beginning of each file to align the nonwords’ offset during the task and to randomize the nonword start and the interval between nonwords.

A second task explored the effect of context, in which the participants were asked to track a phoneme substitution in a word within a sentence (Pittman and Schuett, 2013). The detection task was composed of two lists of 36 four-word-sentences. The original sentences in the lists were composed of simple words from a 3 year-old child’s vocabulary (Bleses et al., 2008a, 2008b) to guarantee that the words included would be well known by the participants. The sentences were evaluated as highly meaningful by a group of 25 native Danish speakers, in a pre-study conducted by our group. For half of the sentences in each list, the first phoneme of the second word was substituted for another phoneme with similar phonotactic probability (e.g., Hunden finder altid maden [the dog always finds the food] - > Hunden sinder altid maden, an equivalent example in English would be ‘Dad buys new shirts’ - > ‘Dad fuys new shirts’, from Pittman and Schuett (2013)). We were careful to choose phonemes that would generate a nonword when replacing the original phoneme, which was confirmed by a group of 14 native speakers who listened to the generated nonwords in isolation and were asked to write the first real word it would remind them of (less than 50% of the participants could point to same original or other real word). Due to this requirement, the phonemes selected had contrasts in one or more production features with the original phoneme, which potentially added cues that may have aided detection in the sentence context. Moreover, to avoid that the second word in the sentence could be predicted by the sentence context, the same group of 25 native speakers were asked to complete the sentences where the target word was missing. Only the sentences with less than 10% of participants filling in the same real word (defined as low cloze probability in Kutas and Hillyard, 1984) were included in the lists.

The final 72 sentences, half with embedded nonwords, were recorded by a female native Danish speaker with an accent from the Danish capital region. She was instructed to pronounce the sentences in a natural prosody but in a slow speaking pace. All the recordings were normalized by RMS and silence was added in the beginning of each file to randomize the sentences’ start.

In order to reduce the acoustic features, challenging the detection and discrimination of phonological contrasts (Stenfelt and Rönnberg, 2009), one list was randomly vocoded for each participant. The vocoding process includes dividing the speech signal into frequency bands, extracting the amplitude envelope for each band, and using it to modulate a noise band, resynthesizing the bands to create a new audio file. For this study, the vocoded versions of the stimuli were generated using the software Praat (Boersma and Weenink, 1992) and the open-source code provided by Winn (2021) (version 45). An 8-channel vocoder was used, with flat-spectrum noise-carrier, and corner frequencies set between 0.2-8 kHz. This number of bands was chosen to add challenges to the transmission of spectral information while approaching the asymptotic speech-recognition performance in quiet (Dorman et al., 1997; Friesen et al., 2001; Xu et al., 2005).

Considering that a participant’s inability to recognize real words in the vocoded condition could influence their performance on nonword detection, vocoded word recognition scores were also calculated. The first list of the clinical test Dantale I (Elberling et al., 1989) in silence was vocoded using the method described above. Participants were presented with the recorded monosyllabic words in isolation and were asked to repeat them aloud. The participant’s response was recorded and transcribed, for offline scoring.

For both phonological discrimination and the detection tasks, accuracy for ‘yes/no’ responses was recorded. Analysis was conducted in terms of signal detection theory (Macmillan and Creelman, 2005). The ‘signal’ in the stimulus was defined as the presence of a phonological contrast, namely, the presence of a nonword in the sentence or a phonological substitution in the second token of the nonword-pair. Responses were classified as ‘hits’: correct responses when the signal was present, ‘misses’: incorrect responses when the signal was present, ‘correct rejections’: correct responses when the signal was absent, and ‘false alarms’: incorrectly reporting the presence of the signal when it was absent. The proportion of correct responses was calculated as the sum of ‘hits’ and ‘correct rejections’, divided by the total number of trials (Macmillan and Creelman, 2005).

The discrimination score (d’) was calculated as measure of the participants’ sensitivity to the presence of a signal (Macmillan and Creelman, 2005). It was estimated by subtracting the z-transformed ‘hit’ rates and ‘false-alarm’ rates. To avoid floor and ceiling effects in d’ calculation, a correction for the extreme values was performed using the log linear approach described by Stanislaw and Todorov (1999), by adding 0.5 to both the number of ‘hits’ and ‘false alarms’ and adding 1 to the number of trials, before calculating the d’ score. Additionally, to analyze a possible response bias toward selecting one of the two options (‘yes’/‘no’), the criterion location was calculated as minus half of the sum of z-transformed ‘hit’ and ‘false-alarms’. A positive criterion value indicates a bias to ‘miss’ the signal although it is present, while negative values represent bias toward accusing the presence of the signal despite its absence (‘false alarms’). Together, d’ and criterion location give a parameter of participants’ strategy in the phonological discrimination and detection tasks.

Pupil data were segmented by trial, and data were analyzed from the eye with best overall confidence during the task, calculated by the percentage of data points over 0.85 of confidence, as reported by the equipment software. The data were cleaned of blinks and artifacts by detecting dilation speed outliers with the method described by Kret and Sjak-Shie (2019) and excluding the flagged data points with a backward and forward margin of 50 ms. Data reconstruction was done using Piecewise Cubic Hermite Interpolating Polynomial (Pchip) or linear interpolation when the Pchip was not possible (where there were not enough points available before or after the region to interpolate), considered the good reconstruction properties of both methods reported by Dan et al. (2020). Blinks above 500 ms were not reconstructed. The individual data points were downsampled to 30 Hz, as the pupil response latency of is over 200 ms (Winn et al., 2018; Mathôt and Vilotijević, 2022), and smoothed using a moving-average filter of 0.1 s.

Trials with more than 45% interpolated data were excluded from the analysis (Burg et al., 2021; Zhang et al., 2022). Baseline pupil size was calculated per-trial by taking the mean pupil size during the 500 ms right before stimulus onset (Seropian et al., 2022). All subsequent data points in the trial were calculated as the proportional change relative to that baseline pupil size. As a last step, raw and processed data were visually inspected to identify and exclude trials with potential contamination, as artifacts in the baseline estimation period or absolute changes in pupil size over 40% of the baseline (Winn, 2016; Winn et al., 2018).

Subjects with more than 50% of the trials excluded from one task condition, had their results excluded from the analysis in that specific task. This criterion excluded pupillometric data from three subjects in both conditions of the phonological discrimination task only. For the remaining participants and tests, the aggregated trace of the pupil response for correct answers was calculated. Data was extracted regarding the value and time of the maximum pupil size – respectively, peak pupil dilation (PPD) and the peak pupil dilation latency (PPL) – from the time window spamming from the target stimulus onset (the second word) to 1 s after the audio offset. To compare with studies with similar methodology (Wagner et al., 2016; Winn and Teece, 2021), a growth curve analysis (GCA) was carried out, which models the quadratic fit of the pupil curve between the target-stimulus onset and the PPD.

Inferential analysis was conducted in Python 3.9, using ‘SciPy’ (v. 1.7.3) and ‘Statsmodels’ (v. 0.13.2) packages. Normality in distribution was assessed using Shapiro–Wilk, for the subsequent choice of parametric or nonparametric statistical tests described in the results. Paired comparisons were conducted for mean/median comparison of signal detection performance (d’) in vocoded and non-vocoded conditions, and the sequential points in the pupillometric curve in ‘yes’ versus ‘no’ tasks. Logistic regression was used to investigate how GCA parameters (intercept, slope, and quadratic term) could be modeled to determine the type of pair (‘yes’ or ‘no’) identified. Additionally, effects of vocoding in the pupillometry metrics were analyzed using a linear mixed effect model in a matrix of auditory condition (‘vocoded’ or ‘non-vocoded’) and pair type (‘yes’ or ‘no’), with participants attributed as random effects. The inclusion of pair type in the model derives from the assumption that the detection of a phonological contrast in the target word (‘yes’ tasks) would produce a more prominent response in task evoked pupillometry (Kinzuka et al., 2020).

In the non-vocoded condition, pupil parameters were sensitive to the presence or absence of the phonological contrast. In the phonological discrimination task, the PPL for pairs without contrast occurred, on average, 217 ms later than pairs with contrast, while PPD values had no significant difference (Table 1). A logistic regression analysis showed no significant effects on the pupil curve intercept, slope, quadratic term, or their interactions, between pairs with and without contrast, χ2 (7, n = 44) = 42.6, p = 0.67. In the detection task, participants exhibited greater PPD for sentences containing a nonword compared to sentences containing all real words, with no significant difference in PPL (Table 2). The differences in pupil dilation were significant in the interval of 610 ms to 1750 ms after the nonword onset (Figure 5), although the logistic regression analysis did not show any significant differences in the parameters of the fitted curve between trials with and without the phonological contrast, χ2 (7, n = 37) = 33.9, p = 0.45. No differences between trials with or without the phonological contrast were found for the vocoded stimuli.

In contrast to the perceptual results, there was no effect of speech degradation on pupil measures in the phonological discrimination task. This was found when analyzed with a linear mixed effects model that included auditory condition (vocoded or non-vocoded) and pair type (phonological contrast present or absent) as fixed effects, and participant as a random effect (models’ marginal R²s = 0.001 and 0.021, conditional R²s = 0.394 and 0.311, for PPD and PPL, respectively).

For the detection task, a higher PPD (β = 0.014, p = 0.007) was seen as an effect of speech degradation, when considered in a similar condition x pair type linear mixed effects model, with participants as random effects (marginal R² = 0.037, conditional R² = 0.611). There was no effect of vocoding on PPL (β = 0.359, p = 0.102, marginal R² = 0.032, conditional R² = 0.103).

The presence of a phonological contrast did not elicit higher pupil dilation in the phonological discrimination task, as it would be expected in a presence of a variant stimuli (Wagner et al., 2016; Kinzuka et al., 2020). The differences here can be attributed to the demands of the tasks. The simplicity of the forced-choice task might not have elicited sufficient differences in the demand for cognitive processing to capture the effect of the phonological contrast. Additionally, in contrast to previous studies which used words as material for discrimination, in our study the participants could not use lexical information to support the decision regarding the change in the phoneme category. Therefore, their judgment was forced to occur solely at the phonological level. The absence of a significant difference in pupil parameters suggests that pupil dynamics may be more sensitive to higher-level cognitive and language processing, as to lexical categorization (Kamp and Donchin, 2015), rather than lower-level phonological categorization.

Moreover, the pupil response to phonological contrasts may be indistinguishable from the response for the perception of acoustic contrasts. The contrast between two speakers in our paradigm, one in each token in the nonword-pair, was done to ensure that discrimination was occurring at a phonological rather than acoustical level. It is known that different speakers possess a natural variability in multiple acoustic domains as voice-onset-time, vowel formants, consonant intensity, among others (Allen et al., 2003; Christiansen and Henrichsen, 2011). Therefore, identifying two nonwords as the same would require their processing at the phonological level. However, as the pupil dilates for acoustic deviants, such as pure tones and noise varying in frequency (Liao et al., 2016; Selezneva et al., 2021), pupil dilation could also be an index of the processing of the dynamic acoustic characteristics of speech in an effort to solve ambiguity caused by interspeaker variations in phoneme production and boundaries (Lewis and Bidelman, 2020; Winn, 2020; Reese and Reinisch, 2022; Yu, 2022). Such a response to acoustic differences would explain the comparable PPDs recorded for both pairs with and without phonological contrast, since for both types of pairs the acoustic variability was present.

Interestingly, in the phonological discrimination task, PPLs were shorter for pairs with phonological contrast than for pairs without contrast. Koelewijn et al. (2017) describe the PPL as a measure of the speed of cognitive processing, with shorter latencies indicating faster cognitive processing or the need for processing less information. One explanation for our results is that to correctly identify a phonological contrast, the participant would only need to identify the first phoneme of the second word in the pair, but to correctly identify the absence of a contrast required the processing of the whole nonword in a pair. Therefore, a decision could be taken quicker with far less information for pairs with contrast.

The presence of context, in the detection task, led to higher PPD in sentences containing a phonological altered word (nonword). This effect was expected as it had been previously reported by Wagner et al. (2016) and Winn and Teece (2021, 2022). These studies found that substituted and distorted phonemes within words in a sentence lead to steeper pupil dilation. As discussed in Winn and Teece (2022), the presence of sentence context makes it difficult to determine if the higher dilation occurs due to increased cognitive demand for sentence processing introduced by the ambiguous lexical entry, or due to the detection of the phonological contrast. However, the absence of difference in the results of the phonological discrimination task suggests that the pupil response may be more closely linked to the violation of the lexical expectation rather than the phonological contrast.

Remarkably, in our study, participants were not asked to process the whole sentence in any manner (they did not repeat it back, nor derived its meaning). Therefore, it could be expected that after detecting the nonword in the second position of the sentence, the participants’ demand for processing would immediately decrease, which should have resulted in a reduction in the pupil size. Yet, the observed pupil behavior indicates that the whole sentence was processed before the response was given. Despite the different protocols used, these results are consistent with Winn and Teece (2021, 2022), in which participants were asked to repeat the whole sentence back to the experimenter. These findings suggest that, despite being instructed to track individual words in the sentence, listeners may have used the whole sentence context to make decisions regarding the presence or absence of the phonological contrast. As an anecdotal report, during the experiment session, several participants reported attempting to ‘repair’ the nonword or ‘figure out the correct word’.

The results in the vocoded condition support the argument that the pupil response reflects processing at the lexical and sentence level. The high accuracy scores for identifying the presence of a nonword within a sentence shows that participants were able to detect the phonological alteration despite the vocoded speech, indicating that phonological discrimination was occurring at a low-level. However, the lack of difference in the pupil parameters between sentences with and without phonologically modified words suggests that the pupil response captured the increase in cognitive processing required to understand the vocoded sentences, rather than the detection of a phonological contrast. Furthermore, the trend of interpreting nonwords as real words in the performance results suggests that the participants were likely attempting phonological restoration throughout the vocoded experiment. In other words, it is possible that the absence of differences in the pupillary response between stimuli with and without phonological contrast reflects the registration of a different type of response besides the detection of the contrast. The physiological mechanisms underlying pupil dilation are also involved in the process of decision-making (Kafkas and Montaldi, 2018). As such, when decisions require greater cognitive processing and memory demand, pupil size increases. It is important to note that the signal restoration of the vocoded stimuli comes at a cost even for real words (Winn et al., 2015; Balling et al., 2017). This global response, which is related to the processing of the auditory stimulus as a whole, may be more pronounced than the response to the detection of the phonological contrast, thereby masking its signal.

Another possible explanation for the lack of difference in pupil metrics between stimuli with or without phonological contrast is that errors in detecting the contrast may have occurred at different moments in the stimuli presentation. Since participants were not instructed about the possible location of the phonological contrast, it was not possible to track the exact moment when errors occurred. As a result, the effect of the phonological contrast may have been distributed across the time series average (Winn and Teece, 2021), which could not be tracked by our analysis.