Introduction
Dyslexia affects approximately 5–10% of the population and is widely recognized as having a lifelong impact on reading and writing skills (Vellutino et al., Reference Vellutino, Fletcher, Snowling and Scanlon2004). Dyslexia is not attributed to low intelligence or poor educational opportunities (Snowling et al., Reference Snowling, Hulme and Nation2020). A core characteristic of dyslexia across languages is difficulties in phonological (speech sound) processing (Stanovich, Nathan & Zolman Reference Stanovich, Nathan and Zolman1988). These difficulties are found at all levels of the linguistic hierarchy, including prosodic patterns (such as stress and intonation), syllable structure, rhyme patterns (e.g., shared endings of syllables or words), and individual phonemes (Goswami, Reference Goswami2018). Phonological difficulties in dyslexia are typically assessed by oral tasks, in which children have to either categorize or manipulate different phonological units such as phonemes, rhymes, or syllables (Ziegler & Goswami, Reference Ziegler and Goswami2005, for review). A profound difficulty in accessing phonological units of all sizes is found consistently in both children and adults with dyslexia, suggesting that the acoustic speech signal is processed atypically by the dyslexic brain. Over time, this results in neural speech-based representations in the mental lexicon of word forms that differ in subtle ways from the representations developed by typically developing children (Keshavarzi et al., Reference Keshavarzi, Mandke, Macfarlane, Parvez, Gabrielczyk, Wilson, Flanagan and Goswami2022; Tan et al., Reference Tan, Chanoine, Cavalli, Anton and Ziegler2022).
One potential perceptual basis for these phonological difficulties lies in atypical categorical perception (CP) of speech sounds. CP allows listeners to distinguish continuous acoustic variations as discrete phoneme categories. For example, the syllables “pa” and “ba” differ in their initial phoneme (/p/ versus /b/). This distinction is primarily based on differences in voice onset time (VOT)—the time interval between the release of the stop consonant and the onset of vocal cord vibration. To produce both /p/ and /b/, airflow is temporarily obstructed by closing the lips (a bilabial closure). The key difference lies in vocal cord activity: for /b/ (a voiced stop), the vocal cords begin vibrating almost immediately upon release, whereas for /p/ (a voiceless stop), there is a longer delay before voicing begins. When the VOT crosses a critical threshold, listeners perceive the change from /p/ to /b/.
Perceptual discrimination experiments show that the phoneme boundary occurs at the same temporal point in a CP task for infants, adults, and a range of animal species including chinchillas, birds and insects (Eimas, Reference Eimas1975; Kuhl, Reference Kuhl2004; Liberman, et al., Reference Liberman, Harris, Hoffman and Grifith1957). The animal data in particular suggest that the phonetic repertoires of the languages of the world are thus capitalizing on natural auditory discontinuities when encoding a continuous signal (see also McMurray, Reference McMurray2022). There are also effects of linguistic experience in human CP, as (for example) temporal points of obstruction that occur within a category boundary are also perceived by younger infants (Kuhl, Reference Kuhl2004). This “allophonic” perceptual ability is thought to decline in the first year of life (Kuhl, Reference Kuhl2010), a phenomenon known as perceptual narrowing. However, some CP studies have suggested that pre-school children and also school-aged children with dyslexia retain allophonic perception (Noordenbos & Serniclaes, Reference Noordenbos and Serniclaes2015). Children with dyslexia can show an increased ability to discriminate within-category speech sounds when compared to typically developing children (e.g., Serniclaes et al., Reference Serniclaes, Van Heghe, Mousty, Carré and Sprenger-Charolles2004; Bogliotti et al., Reference Bogliotti, Serniclaes, Messaoud-Galusi and Sprenger-Charolles2008; Serniclaes & Seck, Reference Serniclaes and Seck2018). This ability is unexpected given that perceptual narrowing should be occurring. Further, it suggests that the linguistic information about phonemes carried by the continuously varying physical acoustic waveform of speech is represented differently in the brains of children with dyslexia (temporal sampling (TS) theory, Goswami, Reference Goswami2011). Critically, such disruptions might have downstream consequences for phonological awareness—the ability to consciously reflect on and manipulate phonological units such as syllables, rhymes, and phonemes. If children with dyslexia experience difficulties in consistently categorizing speech sounds due in part to atypical CP, it follows that they would struggle with tasks that require them to segment, blend, or manipulate phonological information. Thus, atypical CP might be seen as a perceptual foundation underlying the well-documented phonological awareness deficits in dyslexia.
The studies showing allophonic perception in dyslexia indicate greater perceptual sensitivity to fine-grained speech information, thereby appearing to contrast with other perceptual studies in which children with dyslexia show reduced performance in the classic CP task (Breier et al., Reference Breier, Fletcher, Denton and Gray2004, VOT; Cheung et al., Reference Cheung, Chung, Wong, McBride-Chang, Penney and Ho2009, VOT and tone; Maasen et al., Reference Maasen, Groenen and Cru2001, VOT; Noordenbos et al., Reference Noordenbos, Segers, Serniclaes, Mitterer and Verhoeven2012a, place of articulation; O’Brien et al., Reference O’Brien, McCloy, Kubota and Yeatman2018, place of articulation and spectral envelope; Reed, Reference Reed1989, place of articulation). Not all child CP studies find impaired CP in dyslexia, however (Blomert & Mitterer, Reference Blomert and Mitterer2004, place of articulation; Messaoud-Galusi et al., Reference Messaoud-Galusi, Hazan and Rosen2011, VOT), complicating the picture further. For example, Messaoud-Galusi et al. (Reference Messaoud-Galusi, Hazan and Rosen2011) found no consistent difference in CP performance between dyslexic and typically developing children (age and reading level matched) when using a VOT continuum. Rather, the deficit might appear under specific experimental conditions (e.g., quiet environments, task-specific demands on attention or memory such as task-specific differences in discrimination versus identification tasks). Supporting this interpretation, Noordenbos and Serniclaes (Reference Noordenbos and Serniclaes2015) conducted a meta-analysis of 36 CP studies and concluded that dyslexia involves both reduced between-category discrimination and enhanced within-category discrimination, consistent with the notion of “allophonic perception.” Their analysis yielded a large mean effect size (d = 0.86 for discrimination; d = 0.66 for identification) when comparing dyslexic children to age-matched typically developing (TD) control children, but the effect was substantially smaller (d = 0.32) when compared to reading-level (RL) matched controls, and too few discrimination studies existed to calculate effect sizes for discrimination studies.
In summary, while the presence of allophonic perception points to heightened sensitivity to subphonemic variation, CP findings in dyslexia are far from uniform. The variability across studies suggests that CP deficits may be present only in specific subgroups and may only emerge under particular cognitive or perceptual demands or may reflect reading-related delays rather than categorical impairments. This highlights the need for more nuanced research that considers task design, developmental stage, and individual differences within the dyslexic population.
From the perspective of understanding causal factors in reading development, RLmatch studies are extremely important (Goswami, Reference Goswami2003). It is well established in the behavioral literature that there are developmental phonological “grain size” effects across languages. Phonological awareness in children develops from an awareness of larger units (larger grain sizes) such as syllable stress patterns, syllables, and rhymes to an awareness of phonemes (smaller grain size), the latter developing as alphabetic reading is taught (Ziegler & Goswami, Reference Ziegler and Goswami2005). Before literacy is taught, pre-reading children (and illiterate adults) perform poorly on phoneme-level tasks (e.g., “What is the second sound in the word ‘train’?”). It is therefore quite possible that some of the variation in the CP literature depends on how much reading experience the children with dyslexia in the different studies have received. As all children learn about discrete phoneme categories as part of learning to read an alphabetic script, individual differences in exposure to reading may in themselves cause variability in demonstrating CP.
CP studies in preschool children, particularly those AR for dyslexia, are therefore theoretically extremely important. Horlyck et al. (Reference Horlyck, Reid and Burnham2012) studied both native and non-native CP in TD Australian 5- and 6-year-olds and reported that children with more schooling experience showed reduced abilities to discriminate allophonic variations in VOT native contrasts. This could suggest that learning the alphabet at school was already eroding allophonic perception for the latter group. Noordenbos et al. (Reference Noordenbos, Segers, Serniclaes, Mitterer and Verhoeven2012a) studied Dutch kindergarten children AR for dyslexia and found both weaker discrimination of acoustic differences between phonemic categories and better discrimination of acoustic differences within phonemic categories (allophonic perception). After 6 months of reading instruction in the first grade, the “CP deficit” was no longer present for these children in behavioral tasks, although allophonic perception was found to be preserved in the AR children using neural measurements (mismatch negativity [MMN], discussed further below, (Noordenbos et al., Reference Noordenbos, Segers, Serniclaes, Mitterer and Verhoeven2012b). Other preschool AR studies have focused on the classic CP task. Both Boets et al. (Reference Boets, Vandermosten, Poelmans, Luts, Wouters and Ghesquiere2011, 5-year-olds) and Gerrits and de Bree (Reference Gerrits and de Bree2009, 3-year-olds) reported diminished across-category perception of stop consonants in AR preschoolers, but did not assess allophonic perception. It is also important to note that phonemes in natural speech almost never occur in isolation in the physical signal. Accordingly, behavioral CP tasks may draw on additional cognitive skills, such as working memory, attention, executive functioning, and verbal processing speed, which may also differ between children with dyslexia and TD control children, or between preschool children AR for dyslexia and TD preschoolers.
One method for examining CP in dyslexia that controls for these other cognitive factors is to use neural measurements. The most popular neural measure for phonetic discrimination is the MMN, a frontocentral neural response typically elicited in an oddball paradigm. In oddball paradigms, participants are presented with a series of standard stimuli interspersed with one or more deviant stimuli that differ acoustically from the standard, eliciting a negative deflection in the event-related potential (ERP) waveform around 150–250 ms after the onset of the deviant stimulus (Näätänen et al., Reference Näätänen, Paavilainen, Rinne and Alho2007). The MMN serves as an index of auditory discrimination and phonological processing, reflecting the brain’s ability to automatically detect deviations from expected auditory patterns. In children, the MMN can exhibit variability in latency and morphology. Preschool-aged children often demonstrate a delayed and broader peak, with responses occurring between 120 and 400 ms (Maurer et al., Reference Maurer, Bucher, Brem and Brandeis2003; Shafer et al., Reference Shafer, Yu and Datta2010). This variability highlights the ongoing maturation of auditory processing systems in young children, suggesting that their neural mechanisms for phonological discrimination may still be developing. Additionally, a late discriminative negativity (LDN) can be observed at 300–600 ms post-stimulus onset. While the exact function of the LDN remains unclear, it is hypothesized to be associated with higher order cognitive processing (Bishop et al., Reference Bishop, Hardiman and Barry2011) or the reorientation of attention (Horváth et al., Reference Horváth, Czigler, Birkás, Winkler and Gervai2009).
In line with classic behavioral CP findings, several neural studies have reported diminished discrimination of speech contrasts that cross phonetic categories in children with dyslexia. For example, studies using the MMN component have shown reduced neural responses to consonants (e.g., /da/-/ba/ contrast; Schulte-Körne et al., Reference Schulte-Körne, Deimel, Bartling and Remschmidt1998), vowels, and vowel durations (Lovio et al., Reference Lovio, Näätänen and Kujala2010; Männel et al., Reference Männel, Schaadt, Illner, van der Meer and Friederici2017). A meta-analysis of 10 MMN studies conducted with children with dyslexia and TD control children aged 7–12 years (Gu & Bi, Reference Gu and Bi2020) reported a reliable difference in MMN amplitude (effect size d = 0.296). However, as with behavioral studies, findings across neural studies are not uniform. Many of the stimuli used were CV syllables differing in formant transition, VOT, or pitch—manipulations relevant to phonetic categorization, though not always tested in CP paradigms per se. Recent work further illustrates this complexity. Chen (Reference Chen2022), using MMN design with 20-month-old Dutch learning toddlers with and without a familial risk for dyslexia (AR vs NAR), tested sensitivity to a native vowel contrast [ɣɪp] and [ɣɪp] (giep and gip) under both single-speaker and multi-speaker conditions. Contrary to the prediction that AR toddlers would show allophonic-like over-sensitivity to within-category acoustic variation in the multi-speaker condition, both groups exhibited comparable MMN amplitudes across conditions. Yet, AR toddlers showed a delayed MMN peak latency, suggesting differences not in whether a contrast was detected but how and when it was processed. Taken together, these neural findings mirror the heterogeneity observed in behavioral CP studies: while some children with dyslexia (or at risk) exhibit reduced neural discrimination, others show no deficit—or different processing patterns altogether.
The TS theory (Goswami, Reference Goswami2011) provides a relevant framework for understanding the neural basis of allophonic perception in dyslexia. According to this theory, children with dyslexia develop atypical phonological representations of speech because of sensory/neural difficulties in processing the continuously varying speech amplitude envelope at frequencies <10 Hz (Goswami, Reference Goswami2011). The amplitude envelope is the slow-varying energy profile of the speech signal, and it carries crucial low-frequency acoustic information that is used for linguistic decoding (Greenberg et al., Reference Greenberg, Arai and Grant2006). Consistent with TS theory, children with dyslexia exhibit neural impairments in low-frequency speech envelope encoding (Di Liberto et al., Reference Di Liberto, Peter, Kalashnikova, Goswami, Burnham and Lalor2018; Keshavarzi et al., Reference Keshavarzi, Mandke, Macfarlane, Parvez, Gabrielczyk, Wilson, Flanagan and Goswami2022; Power et al., Reference Power, Colling, Mead, Barnes and Goswami2016). Low-frequency envelope information <10 Hz carries important phonetic information, which is known to be used during continuous speech listening by infants when setting up their phonetic inventories (Di Liberto et al., Reference Di Liberto, Attaheri, Cantisani, Reilly, Ní Choisdealbha, Rocha, Brusini and Goswami2023). Accordingly, impairments in low-frequency speech envelope encoding may cause phonetic information to be encoded differently in dyslexia, with possibly greater reliance on neural sampling of the higher frequency envelope information that also carries phonetic information (Giraud & Poeppel, Reference Giraud and Poeppel2012; Goswami, Reference Goswami2011, Reference Goswami2015, Reference Goswami2020). TS theory would thus expect that children with dyslexia may exhibit allophonic perception. TS theory attributes the allophonic mode to the documented difficulties in extracting linguistic information via sampling low-frequency speech information, which is compensated for by enhanced sensitivity to high-frequency speech information. Consistent with this prediction, it has recently been shown that the perception of temporal fine structure is not impaired in children with dyslexia (Flanagan et al., Reference Flanagan, Moore, Wilson, Gabrielczyk, Macfarlane, Mandke and Goswami2025). In a similar vein, it has also been proposed that there is “phonemic oversampling” in dyslexia, with enhanced encoding of high-frequency information >40 Hz (Lehongre et al., Reference Lehongre, Morillon, Giraud and Ramus2013). Such oversampling could also lead to enhanced allophonic perception. However, relevant neural data, particularly from children with dyslexia, are still rare. Due to the confounding effects of reading experience, pre-reading children at family risk for dyslexia provide a particularly important participant pool for testing these theories. Pre-reading AR children provided data for the current study.
Here, we investigate the presence and extent of CP deficits in children aged 4–5 years with a family risk for dyslexia (at risk, AR) and compare them to children not at risk (NAR). Our study differs from previous MMN research in two key ways. First, we test preschool-aged children—prior to formal reading instruction—whereas most MMN studies focus on older children already showing reading difficulties. This allows us to examine early neural markers of phonological processing before reading experience can shape performance yet after acquisition of the phonological system of the native language is well under way (Werker, Reference Werker2024). Second, unlike prior studies that typically use across-category phonemic contrasts (e.g., vowels or place of articulation), we specifically test both within-category and across-category contrasts using theVOT. The /ba/-/pa/ contrast is particularly suited for investigating CP, as it allows us to probe sensitivity to phonemic boundaries in a tightly controlled acoustic continuum. This paradigm is also central to dyslexia research, where impaired sensitivity to VOT contrasts has previously been reported (e.g., Breier et al., Reference Breier, Fletcher, Denton and Gray2004). By including both within- and across-category deviants, we can assess whether AR children exhibit an allophonic processing style—responding to fine-grained acoustic differences regardless of category—while NAR children are expected to respond only to phonemic (across-category) changes. Specifically, we hypothesize that AR children will show mismatch responses to both within- and across-category contrasts, whereas NAR children are hypothesized to respond only to across-category contrasts, reflecting typical CP.
Methods
Participants
In total, 26 4.5-year-old Australian English-learning children from [anonymized for review] longitudinal cohort participated in this study. Data sets from two children were excluded from the analysis due to more than 10% of bad channels (see EEG processing and analysis). The final data set included 24 4.5-year-old children (M age = 53.73 months; SD = 2.86; 14 females, 10 males). Of the 24 children, 11 children were at risk for dyslexia (AR group) (M age = 52.57 months; SD = 3.12 months; 9 females, 2 males), and 13 children (M age = 55.11 months; SD = 1.68, 5 females, 8 males) were not-at-risk (NAR control group). We found no differences between the MMN processing and gender (see supplementary materials).
The children included in this study were part of the longitudinal Seeds of Literacy project. This project followed the development of infants at and not at family risk for dyslexia from 5 months to 5 years of age, including assessments of auditory processing, phonological and lexical abilities, emerging phonological awareness and pre-reading skills in preschool, as well as the quality of children’s early language input (Kalashnikova et al., Reference Kalashnikova, Goswami and Burnham2018, Reference Kalashnikova, Goswami and Burnham2019b, Reference Kalashnikova, Goswami and Burnham2019a, Reference Kalashnikova, Goswami and Burnham2020). Children were growing up in Australia, acquiring English as their primary language, and they were all born full term, without additional health complications, and did not have any other reported developmental concerns. The sample included in this study was selected based on families’ ability to visit the lab when their children were 4.5 years and children’s willingness to participate in the EEG task.
Dyslexia has a strong genetic component, with familial aggregation studies indicating that children with a first-degree relative with dyslexia are at significantly increased risk themselves (Pennington & Olson, Reference Pennington and Olson2005). Therefore, children’s assignment to the AR or NAR group in the Seeds of Literacy project was based on a comprehensive screening battery completed by their parents when the children were 5 months of age, at the start of the project. Children were included in the AR group if (1) one parent was diagnosed with dyslexia and/or (2) if at least one parent scored 1.5 SD or more below the standardized mean in a measure of word or non-word reading and in at least two of the following tests—oral reading (accuracy, fluency, and rate), spelling, rapid picture naming (RAN), and digit span, (2) indicated history of experiencing reading difficulties in childhood, and (3) obtained an average score (within .5 SD from the standardized mean) on a measure of nonverbal IQ (only one child was included whose parents scored within the normal range, but who had two older siblings diagnosed with dyslexia). A child was allocated to the NAR group if both their parents obtained scores within .5 SD from the average on all screening tests (see Kalashnikova et al., Reference Kalashnikova, Goswami and Burnham2018 for detailed information about parental screening and Supplementary materials Table S1 for parental scores).
Children completed a battery of standardized tasks assessing various cognitive and linguistic skills. These included measures of phonological awareness, phonological memory, rapid automatized naming (RAN) for both symbolic and non-symbolic items, sentence repetition (as assessed by the CELF), letter identification, grammatical competence (TROG), receptive vocabulary (PPVT), and verbal and nonverbal intelligence. Group comparisons between AR and NAR children were conducted using independent samples t-tests. The results revealed patterns consistent with expectations: although not all differences reached statistical significance, the direction of effects consistently showed poorer performance in the AR group, with several comparisons yielding moderate to large effect sizes (see Table 1 and see Table S2 in the supplementary materials for the descriptive statistics).
Table 1. Results from the standardized test comparing AR and NAR children
Stimuli
We used the same stimuli as in Burns et al. (Reference Burns, Yoshida, Hill and Werker2007), who demonstrated that by the end of their first year, monolingual English-learning infants already exhibit robust CP of this phoneme contrast. Their findings underscore the utility of the ba-pa paradigm in studying speech perception across various populations. The ba-pa paradigm is ideal because it exploits a well-established, easily measurable contrast in VOT between voiced and voiceless stop consonants (/b/ and /p/), which allows for clear differentiation between phoneme categories. The stimuli consisted of the natural production of the syllables [ba], [pa], and [pha]. A female monolingual English speaker recorded multiple tokens of the syllables, and one token of each production was used for further adjustments of the VOT to match the with- and across category boundaries (see Burns et al., Reference Burns, Yoshida, Hill and Werker2007 for more details). The Standard, [ba], had a VOT of 8 ms; the within-category boundary Deviant, [pa], had a VOT of 28 ms, and the across-category boundary, [pha], had a VOT of 48 ms, see Figure 1. All recordings were digitalized with a sampling rate of 44.1 kHz. The syllable duration for the Standard was 450 ms, 450.7 ms for the Within Deviant, and 450.3ms for the Across Deviant. All stimuli were presented at 75 dB SPL.
Figure 1. Waveforms and spectrograms of the stimuli.
Procedure
Children sat on their parent’s lap/ sat alone, approximately 60 cm away from an LCD screen and listened to the stimuli while watching a silent animated movie and/or an experimenter waving colorful toys. Audio stimuli were presented over loudspeakers on the screen’s left and right sides. Stimuli were presented in a double-deviant oddball paradigm, and the stimuli presentation was controlled using Presentation 16.3 (Neurobehavioral SystemInc., www.neurobs.com), which was run on a PC. The interstimulus interval was a 500 ms. The task included 800 trials, of which 80% (640 trials) were standards, 20% (160 trials) were deviants, and 10% (80 trials) were each deviant type. The high number of standard stimuli (640 out of 800 trials) is essential for eliciting a robust MMN response, as it creates a stable predictive pattern in the listener’s brain. This allows the brain to detect deviations when deviant sounds are presented, amplifying the MMN response by contrasting the unexpected sounds with the predictable standards (e.g., Näätänen et al., Reference Näätänen, Paavilainen, Rinne and Alho2007). Each task began with 20 standards, followed by a mixed presentation of the standards and deviants. The first initial 20 standards and standards immediately following deviants were excluded from the analysis. The presentation of the deviants was pseudorandomized with the constraint that a minimum of two standards and a maximum of eight standards were presented between deviants. The maximum length of the experiment was around 20 minutes.
EEG processing and analysis
The continuous EEG data were recorded using a state-of-the-art 129-channel Hydrocel Geodesic Sensor Net (HCGSN), NetAmps 300 amplifier, and NetStation 4.5.7 software, all provided by EGI Inc. The sampling rate was 1000 Hz, with the reference electrode positioned at Cz. To maintain data quality, electrode impedances were diligently kept below 50 kΩ. The continuous EEG data were saved for subsequent offline analysis.
The EEG analysis was performed using EEGLAB (Delorme & Makeig, Reference Delorme and Makeig2004) FieldTrip (Oostenveld et al., Reference Oostenveld, Fries, Maris and Schoffelen2010) by applying the Maryland analysis of developmental EEG (MADE) pipeline (Debnath et al., Reference Debnath, Buzzell, Morales, Bowers, Leach and Fox2020) and customized scripts in MATLAB2022a. The data were initially downsampled to 250 Hz to facilitate computational efficiency and reduce processing time. To account for potential issues in pediatric EEG data, the outer ring of electrodes on the 128-channel net, which often exhibits poor connections in such data (as detailed in Debnath et al., Reference Debnath, Buzzell, Morales, Bowers, Leach and Fox2020), was removed before proceeding with the data analysis. Specifically, the following electrodes were excluded from the analysis: E17, E38, E43, E44, E48, E49, E113, E114, E119, E120, E121, E125, E126, E127, E128, E56, E63, E68, E73, E81, E88, E94, E99, and E107. This left us with 104 channels for further analysis.
The continuous data were filtered to reduce noise and highlight relevant EEG frequency ranges. A high-pass filter with a cutoff frequency of 0.3 Hz and a low-pass filter with a cutoff frequency of 30 Hz were applied. These filters were implemented using the FIR filter provided within the FIRfilt plugin of EEGLAB (Widmann et al., Reference Widmann, Schröger and Maess2015). The FASTER plugin (Nolan et al., Reference Nolan, Whelan and Reilly2010) of EEGLAB was employed to identify and mitigate the impact of bad channels. The MADE pipeline evaluates three standardized (Z-score transformed) metrics across all electrodes: the Hurst exponent, correlation with other channels, and channel variance. Channels exceeding a threshold value of 3 in any of these metrics were classified as bad channels. On average, 6.20 (SD = 1.58) channels met this criterion in our dataset and were therefore removed. Two participants exhibited more than 10% of bad channels and were excluded from the analysis. A copy of the dataset was created to enhance the data quality and remove artifacted components, and ICA was performed on this copied dataset. The resulting ICA weights were then applied back to the continuous data. Artifacted independent components were identified and removed using the adjusted-ADJUST method, as described by Leach et al. (Reference Leach, Morales, Bowers, Buzzell, Debnath, Beall and Fox2020). The data were segmented into epochs ranging from −100 ms to 600 ms relative to the stimulus onset. A baseline correction was applied by removing the data from −100 ms to 0 ms. To minimize the impact of extreme artifacts, data epochs with voltage values exceeding ±150 μV were rejected from further analysis. Spherical spline interpolation, as implemented in the EEGLAB toolbox, was employed to interpolate channels in cases where more than 10% of channels had to be interpolated; the entire epochs were rejected. On average, AR children had 430.54 (SD = 127.44) artifact-free trials for Standards, 75.15 (SD = 22.62) artifacts-free trials for the Within Deviant, and 74.46 (SD = 21.23) artifact-free trials for the Across Deviant. NAR children had 404.08 (SD = 168.63) artifact-free trials for Standards, 71.17 (SD = 28.76) artifacts-free trials for the Within Deviant, and 71.42 (SD = 28.95) artifact-free trials for the Across Deviant. The final step involved referencing the data by computing an average reference data from all channels.
Statistical analysis
Statistical analyses were performed using MATLAB® version R2022a (The MathWorks Inc, 2022) and the FieldTrip toolbox (Oostenveld et al., Reference Oostenveld, Fries, Maris and Schoffelen2010), as well as R, version 4.1.2 (R Core Team, 2024). To identify relevant time windows and electrode clusters where Standard and Within Deviant and Standard and Across Deviant ERPs significantly differed, nonparametric cluster-based permutation tests (Maris & Oostenveld, Reference Maris and Oostenveld2007, p < .05, α = .05, 1000 permutations, ≥ 3 channels minimum cluster size) were performed for each deviant and assessment point separately. All electrodes except the outer ring (see EEG procedure and analysis) and all time points between 0 and 600 ms were included.
Results
In the AR group, the cluster-based permutation test revealed a difference between the Standard and Within Deviant (p < .05, Cohen’s d = 0.84). This corresponded to a negative cluster in the data. The negative cluster started around 350–600 milliseconds after stimuli onset and was most pronounced over frontal electrodes (see Figure 2, for a distribution across other regions, see S1 for a distribution of the Within Deviants across regions in the supplementary materials). In this time window and cluster, the average amplitude for the Standard was 1.16 μV (SD = 3.20) and for the Within Deviant −1.53 μV (SD = 3.25).
Figure 2. ERP response in AR and NAR children.
Note. Standard and Deviant ERP waveforms of the significant clusters are shown on the left, corresponding topographic information is shown on the right. For the AR children (top) the Within and Across Deviant revealed negative clusters around 300-600 milliseconds after stimuli onset. For the NAR children (bottom), only Across Deviant revealed significant clusters at 150–300 ms. The zero point is at the stimuli onset. Shaded areas indicate the standard error of the mean.
In addition, in the AR group, the cluster-based permutation test revealed a difference between the Standard and Across Deviant (p < .05, Cohen’s d = 1.04). This corresponded to a negative cluster in the data. The negative cluster was around 300–600 milliseconds after stimuli onset and was most pronounced over frontal electrodes (see Figure 2 for distribution of the Across Deviant across other regions, see S2 in the supplementary materials). In this time window and cluster, the average amplitude for the Standard was 1.12 μV (SD = 3.29) and for the Across Deviant −2.02 μV (SD = 2.69).
In the NAR group, the cluster-based permutation test revealed a difference between the Standard and Across Deviant (p < .05, Cohen’s d = 0.58). This corresponded to a negative cluster in the data. The negative cluster started around 150–300 milliseconds after stimuli onset and was most pronounced over posterior electrodes (see Figure 2 for distribution of the Within Deviant and Across Deviant across other regions, see S4 and S5 in the supplementary materials). In this time window and cluster, the average amplitude for the Standard was 0.22 μV (SD = 2.96) and for the Across Deviant −1.54 μV (SD = 3.15).
Since cluster-based permutation testing does not determine whether one deviant elicited a more negative MMN than the other, we conducted an additional analysis to compare the Within Deviant and Across Deviant in the AR group. For statistical analyses, we used the average amplitude of the Standard, Within Deviant and Across Deviant in the respective time window and electrodes as the dependent variable. We conducted the analyses with the statistical software R, version 4.0.4 (R Core Team, Reference Core Team2021) and the lme4 package (Bates et al., Reference Bates, Maechler, Bolker and Walker2015). Plots were created using ggplot2 (Wickham, Reference Wickham and Wickham2016), and post hoc tests were performed using means (Lenth, Reference Lenth2016). P-values were calculated with lmerTest (Kuznetsova et al., Reference Kuznetsova, Brockhoff and Christensen2017), which uses Satterthwaite approximations to degrees of freedom.
We computed a linear mixed effects model with the effect of Condition (Standard, coded as -1 vs the mean of the Within and Across Deviants, both coded as +0.5; and Within Deviant vs Across Deviant, coded as −0.5 and 0.5, respectively). The models included a random intercept by Subject and a random by-subject slope of Condition. All contrast-coding was performed by using the general inverse (Schad, Vasishth, Hohenstein & Kliegl, Reference Schad, Vasishth, Hohenstein and Kliegl2020). In these models, the grand-mean data reflects the intercept (lmer(Amplitude∼Condition+ (1 + Condition|Subject). The results show a statistically significant difference between Standards and Deviants, with Deviants eliciting a more negative response than Standards (β (SE) = −2.91 (0.71), t = −4.11, p = .002). However, the difference between the Within and Across Deviant was not significant (β (SE) = −0.24 (0.47), t = −0.52, p = .616), meaning that AR children processed Within and Across Deviants in a similar way.
The MMN to the Across Deviant did not differ significantly between the two groups (AR vs NAR: β (SE) = −0.298 (1.14), t = −0.261, p = .994). None of the behavioral measures correlated with the MMN response for the Within and Across Deviant (see supplementary materials Tables S3 and S4).
In summary, the results revealed that AR children showed an MMN for Within and Across Deviants, and there was no difference between them. In contrast, the NAR children showed an MMN for the Across Deviant stimuli only.
Discussion
In the current study, we used a neural approach to investigate CP in children aged 4–5 years with a family risk for dyslexia (AR group), comparing their neural responses to children without such a risk (NAR group). Prior CP research with at-risk populations has yielded mixed results, particularly concerning the preservation of allophonic perception, which is the ability to perceive within-category phonetic contrasts (Boets et al., Reference Boets, Vandermosten, Poelmans, Luts, Wouters and Ghesquiere2011; Chen, Reference Chen2022; Gerrits & de Bree, Reference Gerrits and de Bree2009). Our study adds to the current literature in several important ways. First, we examined both within-category and across-category phoneme discrimination in a single MMN paradigm using a well-studied VOT contrast (pa–ba), which allows for a precise manipulation of phonemic category boundaries and acoustic similarity. Second, we focused on pre-reading children, enabling us to isolate effects related to familial risk independent of reading instruction or experience. Our MMN data, which are robust against group differences in attention or cognitive strategies, showed that the AR children demonstrated MMN responses to both within-category and across-category phonetic contrasts, whereas NAR children exhibited MMN responses to across-category contrasts only. Accordingly, only the AR group demonstrated enhanced within-category perception (allophonic perception). These results provide neural evidence that speech processing differs in AR children prior to formal literacy instruction. Preserved allophonic perception in pre-school children at family risk for dyslexia is consistent with earlier findings reported for school-aged children with dyslexia by Serniclaes and colleagues (Bogliotti et al., Reference Bogliotti, Serniclaes, Messaoud-Galusi and Sprenger-Charolles2008; Serniclaes et al., Reference Serniclaes, Van Heghe, Mousty, Carré and Sprenger-Charolles2004, Reference Serniclaes, López-Zamora, Bordoy and Luque2021; Serniclaes & Seck, Reference Serniclaes and Seck2018). Consistent with TS theory, the data suggest that the physical acoustic waveform of speech is represented differently in the brains of children at risk for dyslexia. Impaired representation at the levels of syllables and prosody may be compensated for by enhanced processing of fine-grained acoustic differences such as within-category acoustic changes that may specify a phoneme in other languages (Goswami, Reference Goswami2011; Lehongre et al., Reference Lehongre, Morillon, Giraud and Ramus2013).
Studies of pre-reading children are also important with respect to phonological development for another reason. It is possible that some of the variation in the prior behavioral CP literature depends on the degree of reading experience that the children with dyslexia have received. Prior to the onset of learning to read, phoneme awareness in oral tasks is largely limited to awareness of single phonemes in the onset position in syllables (pre-readers can identify the /p/ in “pat”, but not the /p/ in “lip”; Ziegler & Goswami, Reference Ziegler and Goswami2005, for review). Children learn about discrete phoneme categories as part of learning to read an alphabetic script, and illiterate adults often cannot perform phoneme awareness tasks (Araújo et al., Reference Araújo, Flanagan, Castro-Caldas and Goswami2018; Morais et al., Reference Morais, Cary, Alegria and Bertelson1979). By studying pre-reading children, the impact of reading experience on CP is automatically controlled. However, prior behavioral CP studies with pre-reading AR children have not sought evidence for allophonic perception (Boets et al., Reference Boets, Vandermosten, Poelmans, Luts, Wouters and Ghesquiere2011; Gerrits & de Bree, Reference Gerrits and de Bree2009).
Interestingly, the MMN responses in AR children documented here occurred in a later time window (300–600 ms) compared to the typical window observed in NAR children (150–300 ms). This finding contrasts with previous research indicating that children with dyslexia or AR for dyslexia exhibit attenuated late MMNs compared to their typically developing peers (Alonso-Búa et al., Reference Alonso-Búa, Díaz and Ferraces2006; Maurer et al., Reference Maurer, Bucher, Brem and Brandeis2003; Neuhoff et al., Reference Neuhoff, Bruder, Bartling, Warnke, Remschmidt, Müller-Myhsok and Schulte-Körne2012). Notably, the absence of an early MMN has also been documented for other speech contrasts in children with specific language impairment, where it has been interpreted as indicative of a less automatic or slowed discrimination process (Shafer et al., Reference Shafer, Morr, Datta, Kurtzberg and Schwartz2005; Uwer et al., Reference Uwer, Albrecht and Von Suchodoletz2002). Given that we observe a similar absence of the early MMN component, this delayed response in AR children could also be indicative of a protracted maturation process within their neural systems responsible for phonological and auditory processing, instead of a different mode of auditory processing.
Typically, the early MMN (150–300 ms) reflects the automatic and efficient categorization of phonemes, signaling mature phonological processing mechanisms (Näätänen et al., Reference Näätänen, Paavilainen, Rinne and Alho2007). However, the later MMN responses observed in AR children, also referred to as LDN, suggest that their brains are either still undergoing development in this domain, or are using compensatory strategies to process speech contrasts. Some data suggest that the LDN decreases during childhood, potentially indicating a shift toward more mature and efficient neural processing of phonemic contrasts by adults (Cheour et al., Reference Cheour, Korpilahti, Martynova and Lang2001). Our results suggest either that the maturation process for phonological processing in AR children may be prolonged or that phonological processing depends on different neural acoustic processes. The presence of the late MMN in AR children, absent in the NAR children, is more aligned with developmental differences regarding their auditory processing capabilities.
The current study has some important limitations. First, the sample size is small, and replication with a larger pre-reader sample is required. However, it is interesting to note that around 40% of the current sample were also tested using a ba/wa discrimination as 20-month-old infants, where they exhibited enhanced perception of formant transitions (frequency rise time) compared to TD infants (Götz et al., Reference Götz, Peter, Kalashnikova, Burnham and Goswami2024). Second, VOT is only one acoustic measure of CP, and it is reliant on temporal processing. Different results may be found if CP depended on nontemporal acoustic cues. Nevertheless, our data suggest that a difference in VOT processing and the resulting differences in CP could be early markers of the phonological processing differences that later manifest in reading difficulties. Third, children with a familial risk factor for dyslexia might not necessarily develop the disorder. This necessarily limits the contribution of pre-reader AR data. The current findings should thus be revisited once it is known which children in the SEEDS longitudinal study developed dyslexia. Nevertheless, research with at-risk populations remains critically relevant because it provides insights into the underlying mechanisms that contribute to developmental differences.
In summary, our study revealed significant differences in categorical speech processing between AR and NAR children. The results indicate that AR children demonstrate MMN responses to both within-category and across-category phonetic contrasts, suggesting heightened sensitivity to fine-grained phonetic details within phonemic categories. In contrast, NAR children showed MMN responses only to across-category contrasts, consistent with the typical developmental trajectory for phonemic processing. These findings support the view that allophonic perception is preserved in children with developmental dyslexia (Serniclaes et al., Reference Serniclaes, Van Heghe, Mousty, Carré and Sprenger-Charolles2004).
Supplementary material
For supplementary material accompanying this paper visit https://doi.org/10.1017/S0142716425100271
Data availability statement
All data reported here are available upon request to Marina Kalashnikova at m.kalashnikova@bcbl.eu.
Replication package
Pre-processed EEG data and scripts are available https://osf.io/g78xt/
Acknowledgements
We thank Anne Dwyer, Maria Christou-Ergos, Scott O’Loughlin, and Samra Alispahic for their assistance with participant recruitment, data collection, and data analyses. We also thank all the children and their parents for their valuable time and interest in this research.
Financial support
This research was supported by the Australian Research Council grant DP110105123, “The seeds of literacy in infancy: empirical specification of the acoustic determinants of language acquisition Seeds of Literacy,” to the 4th and 5th authors.
Competing interests
The authors declare no conflicts of interest.
Ethical standards
This study was approved by the Western Sydney University Human Research Ethics Committee (approval number H9660).
Open access
