Cortical Tracking of Speech Across Rhythm Classes in English‐, French‐, and Japanese‐Learning 4‐ and 8‐Month‐Old Infants Authors: Varghese Peter, Marina Kalashnikova, Sandrien van Ommen, Reiko Mazuka, Denis Burnham, Thierry Nazzi Published: 2026 ABSTRACT ABSTRACTIt has been proposed that the world's languages fall into three rhythmic classes according to the units that determine their rhythmic structure: stress‐based (e.g., English), syllable‐based (e.g., French), and mora‐based (e.g., Japanese). These rhythmic differences modulate language‐specific processing in adults and young infants. It has also been proposed that the syllable is the default rhythmic unit across languages, more prominent in acquisition and processing. This study investigated the development of these language‐specific versus language‐general effects on the neural processing of speech, focusing on cortical tracking at frequency bands potentially corresponding to three rhythmic units (2.5 Hz/foot, 5 Hz/syllable, 10 Hz/mora). We assessed cortical tracking in 119 infants belonging to one of three native‐language groups: English, French, and Japanese, and two age groups: 4 and 8 months. Infants heard CVN (consonant‐vowel‐nasal) syllable strings synthesized using a native‐language voice in their respective native language, a non‐native voice (Polish), and a non‐speech (vocoded) version of the non‐native voice stimuli. In all conditions, the stimuli consisted of 100‐ms morae, combining into 200‐ms syllables, combining into 400‐ms bisyllabic feet. In all three language groups and at both ages, results showed (1) efficient cortical tracking at 5 and 10 Hz, but inconsistent tracking at 2.5 Hz; (2) stronger tracking at 5 Hz, and (3) differential tracking of speech and non‐speech stimuli. These findings establish that language‐general rhythmic processing guides cortical tracking, that the cross‐linguistic syllabic level is dominant until at least 8 months, and that speech stimuli are processed differently than non‐speech stimuli by at least 4 months.Keywords: cortical tracking, crosslinguistic, English/French/Japanese, infancy, linguistic rhythm, speech/nonspeech Introduction Human languages differ in many aspects, one of the most fundamental being the global rhythm of utterances in a language. At the linguistic level, three broad yet distinct classes of language rhythm have been proposed based on different rhythmic units: the stress‐based class, for example, English, based on the stress unit (the trochaic foot) which occurs at a rate of around 2–2.5 Hz; the syllable‐based class, for example, French, based on the syllabic unit with occurs at around 5 Hz; and the mora‐based class, for example, Japanese, based on the (sub‐syllabic) moraic unit, which occurs at around 8–10 Hz (Abercrombie 1967; Pike 1945). The three rhythmic units are linked hierarchically—stress units consist of syllables, which themselves consist of morae. Despite this hierarchy, it has been suggested that the syllable might be the basic, default rhythmic unit because syllables correspond to the cyclical opening of the jaw to produce consonant‐vowel alternations (MacNeilage 1998). This is supported by the finding of a peak in the envelope spectrum of speech stimuli from nine languages with typologically different rhythmic structures at around 4.3–5.4 Hz (Ding et al. 2017), although some cross‐linguistic differences in different rhythm measures have been found (Varnet et al. 2017). These rhythmic classes appear to have psychological reality, both in adulthood and in infancy. Initial evidence for this comes from studies investigating lexical access in adults that have established language‐specific procedures in the use of rhythmic cues. In processing spoken language, native French language listeners use syllables to access words (Mehler et al. 1981), even when listening to English words (Cutler et al. 1986), but not mora information (Otake et al. 1993); native English language listeners use stress (Cutler and Norris 1988) but not syllable information (Cutler et al. 1986) to access words; and native Japanese language listeners use mora but not syllable information to access words (Otake et al. 1993). Given such cross‐linguistic differences, indeed preferences, in adults, how and when do these differences develop ontologically? Developmental studies have explored infants’ ability to discriminate between languages, establishing some discrimination abilities from birth, together with listening preferences for sentences spoken in the native versus a foreign language (Mehler et al. 1988; Moon et al. 1993). Further studies have established that language discrimination at birth is based on rhythmic classes, rather than on the particular language per se. Thus, newborn infants discriminate between languages from different rhythmic classes—stress‐ versus syllable‐ versus mora‐based (Byers‐Heinlein et al. 2010; May et al. 2011; Nazzi et al. 1998; Ortiz‐Barajas et al. 2023; Ramus et al. 2000; Sato, Hirabayashi, et al. 2012), but not between languages from within these rhythmic classes (Nazzi et al. 1998; Ortiz‐Barajas et al. 2023). However, later, by around 4 months, infants can also discriminate between some languages from within the same rhythmic class (e.g., Bosch and Sebastian‐Galles 1997, 2001; Nazzi et al. 2000). Differences in gamma‐band activity are also found to signal same‐rhythm language discrimination in infants after 6, but not 3, months of speech exposure (hence full‐term 6‐month‐olds and preterm 9‐month‐olds, Peña, Pittaluga and Mehler 2010). So, it has been proposed that the cross‐linguistic differences in adults’ language processing and lexical access develop out of, and are bootstrapped by, newborns’ sensitivity to rhythm, and specifically, the rhythmic class of the language they hear around them. To investigate this possibility, several studies have explored word‐form segmentation in infancy. This ability is crucial for word learning, given that very few words are spoken in isolation, and that most words only appear in sentence contexts. Research across various languages shows that word segmentation abilities emerge by 8–10 months (e.g., English: Jusczyk and Aslin 1995; Dutch: Houston et al. 2000; German: Höhle and Weissenborn 2003; Parisian French: Nazzi et al. 2006; Canadian French: Polka and Sundara 2012; Spanish: Bosch et al. 2013; Portuguese: Butler and Frota 2018), with studies occasionally demonstrating word form segmentation even at younger ages (by 4–6 months: Bosch et al. 2013; Berdasco‐Muñoz et al. 2018; Butler and Frota 2018). Importantly, there is also evidence for language‐specificity in word‐form segmentation abilities based on rhythmic classes. In stress‐based English, there is a predominance of trochaic (strong‐weak) over iambic (weak‐strong) two‐syllable words. Accordingly, by 7.5 months, US English‐learning infants segment two‐syllable words that follow trochaic but not iambic stress patterns, while by 10.5 months, infants can segment both trochaic and iambic words (Jusczyk et al. 1999). Such a result was also found for 10‐month‐old infants learning Dutch, a language in which two‐syllable words are also predominantly trochaic (Kooijman et al. 2005,2009). Thus, in stress‐based languages, early segmentation is based on the stress unit, and its use develops over age. Correspondingly, studies on syllable‐based languages have found an advantage for the segmentation of the syllable over disyllabic word‐forms in Parisian French between 6 and 12 months (Nazzi et al. 2006; Nishibayashi et al. 2015), while Canadian French‐learning infants use both the syllable and the disyllabic word‐form at 8 months (Polka and Sundara 2012). This supports early segmentation based on the syllable in these two dialects of French. Taken together, these findings are consistent with a modulation of early segmentation abilities by the rhythm of the native language. Moreover, the fact that syllable‐based segmentation has been found earlier (at 4–6 months, Bosch et al. 2013; Berdasco‐Muñoz et al. 2018; Butler and Frota 2018) than stress‐based segmentation (by 7.5 months at the earliest, Jusczyk et al. 1999) might be taken as evidence that the syllable is the default rhythmic unit, more prominent in acquisition and infant/adult processing than the stress and possibly the moraic units. At present, there is no evidence of mora‐based segmentation in Japanese‐learning infants, but the fact that the discrimination of monomoraic from bimoraic syllables is only evident late in development, at 10 months, suggests that the mora may be the least prominent of the three rhythmic units (Sato, Kato, et al. 2012). These findings attest to crosslinguistic modulations of infant and adult speech processing by linguistic rhythm (language‐specific effects), plus a potentially privileged role of the syllabic rhythm (language‐general effects) at the behavioral level. Are such effects also found at the neural level? One mechanism that allows assessment of language‐specific and language‐general effects of the rhythmic properties of the speech signal on speech encoding and comprehension is cortical tracking of speech. Cortical tracking (or neural entrainment) refers to the synchronization of ongoing neural oscillations with external rhythmic stimuli (Ding and Simon 2014). This process plays a crucial role in speech perception, as the brain aligns the neural activity with the temporal structure of speech signals. By doing so, important information in speech arrives at the periods of highest neural sensitivity, enhancing speech intelligibility (Giraud and Poeppel 2012; Kösem et al. 2018). It has been shown that adults’ neural oscillations at different electrophysiological bands are involved in the processing of speech in various time scales: the delta band (∼0.5–4 Hz) encodes prosodic and intonational information that is correlated to the acoustic envelope of speech at a scale of 500–2000 ms (Bröhl and Kayser 2021; Teoh et al. 2019); the theta band (∼4–8 Hz) encodes syllabic information that is conveyed at the time scale of 150–300 ms (Rimmele et al. 2023); and the gamma band (>30 Hz) encodes finer phonetic cues such as formant transitions (Lizarazu et al. 2019). Developmentally, there is growing evidence for efficient cortical tracking of speech during infants’ first months of life. For instance, Attaheri et al. (2022a, 2022b) measured cortical tracking of nursery rhymes in English‐learning infants at 4, 7, and 11 months and found reliable tracking at the delta and theta bands across the three ages. Theta‐ but not delta‐band tracking efficiency increased with age, consistent with the maturational trajectory of infant electrophysiological activity, dominated by lower frequencies in the first months with faster frequencies emerging over the first year of life (Menn et al. 2023). Interestingly, individual differences in cortical tracking efficiency have been shown to significantly correlate with infants’ concurrent and later vocabulary size, confirming that this mechanism plays a significant role in the successful encoding of the speech signal and in language acquisition (Ní Choisdealbha et al. 2023; Menn, Ward, et al. 2022; Nguyen et al. 2023). Critically, already in the first months of life, infants’ cortical tracking efficiency is modulated by the rhythmic properties of the speech signal. Kalashnikova et al. (2018) found evidence for more efficient cortical tracking of naturally‐produced infant‐ than adult‐directed speech in 7‐month‐old infants acquiring Australian English (see Menn, Michel, et al. 2022 for similar results in 9‐month‐old Dutch‐learning infants). This effect was attributed to the prosodic properties of infant‐directed speech, particularly increased rhythmic regularity and higher energy concentration in the delta band (as opposed to more prominent amplitude modulations in the theta band in adult‐directed speech, Leong et al. 2017). Although there is no crosslinguistic evidence for language‐specific effects on cortical tracking in infants, two studies suggest that significant differences in the tracking of rhythmically‐different languages are already present at birth. Ortiz Barajas and colleagues investigated envelope tracking in French‐learning infants, showing that the newborn brain tracks the amplitude and phase of sentences in both the native language, a nonnative language of the same type of rhythm (Spanish), and a nonnative language of a different type of rhythm (English), attesting language‐general processing; in contrast, at 6 months, amplitude tracking has become language‐specific, while phase tracking remains language‐general (Ortiz Barajas et al. 2021). Moreover, delta and theta neural oscillations in the newborn brain were found to differ for rhythmically different languages, which could provide a basis for newborn's discrimination of linguistic rhythms (Ortiz Barajas et al. 2023). This raises the question of whether cortical tracking would be modulated by the rhythmic properties of the infant's native language. A recent comparative EEG study (Peter et al. 2022) addressed this question in adults. Peter et al. (2022) investigated cortical processing of language rhythms at the foot (2.5 Hz), syllable (5 Hz), and mora (10 Hz) levels in listeners of stress‐, syllable‐, and mora‐based languages by Australian English‐, French‐, and Japanese‐speaking adults. All three groups listened to the same continuous synthesized speech, made of the concatenation of 200‐ms‐long CVN (consonant‐vowel‐nasal) syllables phonotactically legal in all three languages, and synthesized with an English, a French, and a Japanese voice. In these stimuli, the envelope amplitude at the syllable level (200 ms CVNs) was more salient than at the mora level (100 ms CVs and Ns), while there was no peak evident at the foot level (400 ms CVNCVNs), in line with the fact that both syllables in a “foot” were equally salient (the foot here being a bisyllabic spondee). The reasoning was that if cortical tracking is determined bottom‐up, then it should be stronger at the syllable‐ than at the mora‐level, with no tracking at the foot‐level, in all three groups of adults. In contrast, if cortical tracking is modulated by top‐down knowledge of the rhythmic unit of the native language, then each language group should show a boost in cortical tracking at the level of their respective rhythmic units (2.5 Hz for the foot, 5 Hz for the syllable, 10 Hz for the mora); this should be found even if there is no acoustic marking of the foot level, given previous findings establishing that monotone syllable strings are perceived as sequences of trochaic feet by speakers of stress‐based languages (Bhatara et al. 2013,2016; Kaland et al. 2025). In all three groups, cortical tracking was found at both the syllable and mora levels (but not at the foot level), and was most prominent at the 5 Hz syllable rate. This supports language‐general determinants of cortical tracking, and the proposal that syllabic rhythm is a language‐general level of processing. However, results also revealed cross‐linguistic differences, with enhanced tracking by the French‐speaking adults compared to the other groups at both the 10 Hz and especially the 5 Hz level. Since 5 Hz corresponds to the syllable level, which is the rhythmic unit of French, these latter results also indicate language‐specific effects on cortical tracking. Taken together, the findings support a hybrid model of the determinants of adult cortical tracking, involving both bottom‐up language‐general and top‐down language‐specific factors. Note that the tracking effects found and their crosslinguistic modulation were mostly found at 5 Hz, the syllable level, which might relate to the structure of the stimuli, consisting of CVN syllables. The study further compared processing of speech versus non‐speech stimuli, by presenting the adults with sequences synthesized with Polish properties (non‐native to all participants) plus also vocoded versions of the same sequences. This manipulation revealed stronger tracking at 5 and 10 Hz, stronger tracking by the French‐speaking adults, and stronger tracking for speech than non‐speech in all three groups, further supporting a hybrid model of adult cortical tracking of human speech. The above hybrid language‐general/language‐specific effects raise the question of how cortical tracking is set in place during language acquisition, and when any language‐specific effects might emerge. The present study was designed to address these questions. We used an approach similar to Peter et al. (2022) to test 4‐ and 8‐month‐old infants from Australian English (stress‐based), French (syllable‐based), and Japanese (mora‐based) language backgrounds. These two ages were chosen based on prior behavioral literature showing that rhythmic processing, as attested by preferences for specific rhythmic patterns or segmentation abilities, is still language‐general at 4 months, and that language‐specific rhythmic processing appears to emerge in the second half of the first year, from 6 to 9 months onward (e.g., Höhle et al. 2009; Jusczyk et al. 1993, 1999; Skoruppa et al. 2009; though Berdasco‐Muñoz et al. 2018, might provide some evidence of language specificity in segmentation abilities). First, we assessed infants’ cortical tracking of continuous synthesized speech that followed the phonetic and phonotactic properties of their respective native languages, presenting them with the same stimuli as used in Peter et al. (2022) with adults, synthesized in an English, French, or Japanese voice, and evaluated tracking at 2.5, 5, and 10 Hz. Second, to compare tracking of non‐native speech and of speech versus non‐speech, we compared infants’ cortical tracking of the Peter et al. (2022) speech stimuli synthesized in the non‐native voice (Polish) to its vocoded counterpart (vocoded Polish). Given the adult results that support a hybrid model of cortical tracking, and given previous infant behavioral studies on rhythmic acquisition, we expected changes in the determinants of infant cortical tracking, from language‐general bottom‐up determinants at 4 months, to both language‐general bottom‐up and language‐specific top‐down determinants at 8 months. If this were the case, we would observe a concordance of the acquisition markers at the neural level with those at which rhythmic acquisition has been found in behavioral studies. Accordingly, in all analyses and age groups, we predicted, as found in adults, better cortical tracking at the 5 Hz syllabic level than at the 2.5 and 10 Hz levels, irrespective of language background. In addition, regarding the effect of native language, we predicted no cross‐linguistic differences at 4 months, but some language‐specificity by 8 months which, based on the adult findings, would correspond to stronger tracking of the 5 Hz syllabic level in the French‐learning infants than in the other two groups. For the speech versus non‐speech stimuli, we predicted better tracking for speech over non‐speech over all language groups, but it was less clear to predict whether this effect would increase over age. Indeed, the speech advantage may become stronger with age as infants gain more experience with speech, although it may already be well established at 4 months, given evidence that even newborns prefer speech over non‐speech stimuli (Vouloumanos and Werker 2007). Method Participants The participants were 119 infants across three language groups (English, French, and Japanese) and two age groups (4 and 8 months). There were 16 English‐learning 4‐month‐olds (MAge = 4 months, 20 days; range = 4 months, 10 days to 4 months, 29 days; 9 females, 7 males) and 19 8‐month‐olds (MAge = 8 months, 18 days; range = 8 months, 1 day to 9 months, 0 days; 9 females, 10 males) tested in Sydney, Australia. There were 25 French‐learning 4‐month‐olds (MAge = 4 months, 20 days; range = 4 months, 3 days to 4 months, 29 days; 11 females, 14 males) and 19 8‐month‐olds (MAge = 8 months, 17 days; range = 8 months, 1 day to 8 months, 29 days; 10 females, 9 males) tested in Paris, France. And there were 20 Japanese‐learning 4‐month‐olds (MAge = 4 months, 19 days; range = 4 months, 9 days to 4 months, 28 days; 11 females, 9 males) and 20 8‐month‐olds (MAge = 8 months, 17 days; range = 8 months, 0 days to 8 months, 28 days; 8 females, 12 males) tested in Tokyo, Japan. Thirty‐eight additional infants were tested across the groups, but their data were removed from the analysis due to excessive artifacts in their EEG (see EEG/Data Processing). All infants were born full‐term and did not have any reported vision or hearing impairments. All infants were monolingual, that is, they heard their “native” language for at least 90% of the time in their environment, lived in urban communities, and came mostly from middle/upper‐middle class families (though detailed information was not collected). Community of descent information was not collected as collecting such information is illegal in France. Stimuli The stimuli were identical to those of Peter et al. (2022). These consisted of continuous streams of CVN (consonant‐vowel‐nasal) syllables. Ten consonants (/p/, /t/, /k/, /b/, /d/, /g/, /s/, /z/, /ʃ/, /j/), three vowels (/i/, /a/, /ɔ/), and three nasals (/m/, /n/, /ŋ/) were used to synthesize 14 target syllables with a fundamental frequency of 180 Hz, which were phonotactically possible in English, French, and Japanese. When the first consonant was a stop, that consonant was 20 ms in duration and the following vowel was 80 ms in duration. For fricative and affricate onsets, both consonants and vowels were 50 ms in duration. The nasal was always 100 ms in duration. The speech stimuli were synthesized in MBROLA (Dutoit et al. 1996), using four different voices. The native voices were us1‐English, fr4‐French, jp2‐Japanese, and the non‐native voice was pl1‐Polish (https://github.com/numediart/MBROLA‐voices). A fifth, non‐speech vocoded condition was created from the Polish voice, by extracting the broadband envelope of the syllables between 0 and 120 Hz and modulating white noise in the envelope. The syllables were concatenated into 30‐s sequences in a random order with no time gap between the syllables and with the constraint of no immediate repetition of a given syllable. Ten syllables were repeated 11 times (randomly selected) and the remaining four were repeated 10 times in each sequence. Sound sequences were created dynamically during the experiment in Presentation 18.1 (Neurobehavioral Systems, CA, USA). Ten sound sequences were created for each of the three conditions. Procedure and Apparatus Infants’ neural activity was recorded using 128 channel Hydrocel Geodesic Sensor Net (HCGSN), NetAmps 300 amplifier and Netstation (MagStim EGI Inc) software. The EEG was recorded at a sampling rate of 1000 Hz. The reference electrode was placed at Cz, with impedances kept below 50 kΩ at the beginning of the recording. Infants in each language group were tested in three conditions: native voice (English, French, or Japanese, respectively),1 speech in non‐native voice (Polish) and non‐speech (vocoded Polish). The three conditions were presented in three separate blocks with the order being counterbalanced across participants. Data were collected at the three locations (Sydney, Paris, and Tokyo) using identical instrumentation and procedures. Infants either sat on a high chair or on their caregivers’ lap approximately 1 m away from a loudspeaker. During the EEG recording, infants watched a silent cartoon or an experimenter who silently entertained them with toys (the video and experimenter movements were not rhythmic or synchronized to the stimuli). Caregivers, wearing earplugs and headphones, were instructed not to speak to the infants during the recording. EEG/Data Processing The EEG data were analyzed offline using fieldtrip (version 20210325, Oostenveld et al. 2011) and Letswave 7 (https://github.com/NOCIONS/letswave7/) toolboxes in MATLAB 2020b (MathWorks, Natick, MA, USA). First, electrodes in the outer three rings of the HCGSN were removed (except the right and left mastoids) as they are more prone to artifacts due to movement or muscle activity (Lam‐Cassettari et al. 2021; Lenc et al. 2022), leaving 93 electrodes. The continuous EEG was then bandpass filtered between 0.3 and 30 Hz using a zero‐phase windowed sync finite impulse response filter. The high pass was set at 0.3 Hz to balance the preservation of low frequency neural activity with the removal of large amplitude slow drift and movement related artifacts, which are common in infant EEG recordings (Peter et al. 2026; Urbanec et al. 2024). Although this cut off may attenuate EEG activity meaningfully up to 1 Hz, it is lower than the lowest frequency of interest in the current study (2.5 Hz). The EEG was then divided into epochs between 0 and 30 s relative to sound onset. Trials and electrode channels contaminated by artifacts were identified in the following way. Trials with more than 20 bad channels (amplitude fluctuations exceeding ±100 µV) were removed. For trials with fewer than 20 bad channels, the channels with amplitude fluctuations exceeding ±100 µV were interpolated. Channels that were contaminated for more than 50% of the trials (±100 µV criterion) were also interpolated. A channel was not retained if there were more than two bad neighboring channels. None of the accepted trials from participants had more than two bad neighboring channels, and therefore data from all electrodes were retained. Only participants with at least 50% (i.e., 5 of the 10) accepted trials in each of the three conditions were further analyzed. Epochs were then re‐referenced to the average of the mastoid electrodes (which had clean data for all infants included in the final analyses, and were thus never interpolated) instead of an average reference for two reasons: (1) to increase signal‐to‐noise ratio of neural responses to auditory stimuli at the frontocentral electrodes (Cirelli et al. 2016; Mahajan et al. 2017); and (2) because infant EEG typically have a high proportion of interpolated or rejected channels, which undermines the key assumption of the common average reference that the mean potential across electrodes approximates a neutral reference. Epochs were then averaged in the time domain so that the responses not aligned to the rhythmic input were canceled, thereby increasing the signal to noise ratio. The averaging was performed separately for each participant and each condition. The time‐averaged epochs were then transformed into frequency domain using FFT, producing a spectrum of EEG amplitudes between 0 and 500 Hz with a frequency resolution of 0.025 Hz. The signal to noise ratio was further improved by subtracting, at each 0.025 Hz frequency bin, the average amplitude at the neighboring bins (2nd to 12th on both sides). This process relies on the assumption that, in the absence of a periodic EEG response, the signal amplitude at a given frequency bin should be similar to the signal amplitude of the mean of the surrounding frequency bins. This process also centers the obtained spectra around zero by correcting for the 1/f shape of the background noise (Mouraux, et al. 2011; Retter and Rossion 2016). To avoid electrode selection bias, the spectrum was averaged across all scalp electrodes for each stimulus. This was done because the topography at the frequencies of interest (2.5 Hz, 5 Hz, 10 Hz) were different for different ages and conditions. The significance of the response at the frequencies of interest was calculated using z‐scores. The z‐score at a given frequency was computed as the difference in amplitude between that frequency and the mean of the 20 neighboring frequency bins (10 on either side), divided by the standard deviation of the 20 neighboring bins. The 20 neighboring bins represented a frequency range of 0.50 Hz (0.25 Hz on either side) and excluded the two immediately adjacent frequency bins. In line with previous studies (Quek et al. 2018; Lenc et al. 2022), a z‐score greater than 3.1 was considered significant (p < 0.001, one‐tailed). Determination of EEG Peaks The EEG spectra for the native speech, and for the non‐native speech and non‐speech stimuli for all participant groups are shown in Figures 1 and 2, respectively. Visually, the response at 5 Hz was most prominent followed by the response at 10 Hz. Table 1 shows the z‐scores depicting selective enhancement of brain activity at the frequencies of interest (2.5 Hz, 5 Hz, 10 Hz). The response at 5 Hz was statistically significant for all groups, ages, and stimuli (18 out of 18 conditions). Significant 10 Hz response (15 out of 18 conditions) was seen for all stimuli at 4 months for both the English and Japanese groups, and at both 4 and 8 months for the French group. The presence of 10 Hz response at 8 months for the Japanese and English infants depended on the stimuli, with the vocoded stimuli generating a significant response in both groups, and the non‐native stimuli generating a significant response only in the Japanese group. Interestingly, but consistent with the adult neural responses to similar stimuli (Peter et al. 2022), the response at 2.5 Hz was not significant for any of the groups or ages, with the exception of the 4‐month‐old English infants for speech only (native and non‐native speech stimuli, but not vocoded stimuli). Given the weak and sparse cortical tracking at 2.5 Hz, exploring the effects of age, language, or stimulus type was not justified. Accordingly, all subsequent analyses investigating these effects were conducted using solely the 5 and 10 Hz data. Spectra and topographies of the EEG responses to native stimuli across the language and age groups. Spectra and topographies of the EEG responses to non‐native (Polish) and non‐speech (vocoded Polish) stimuli across the language and age groups. TABLE 1z‐scores computed to determine whether significant EEG responses were elicited at the frequencies of interest (2.5, 5, and 10 Hz) for each stimulus type for each language and age group.GroupAgeStimuli2.5 Hz5 Hz10 HzEnglish4 monthsNative 4.77 * 12.92 * 5.89 * Polish 3.16 * 18.44 * 3.53 * Vocoded−0.72 12.78 * 7.38 * 8 monthsNative0.06 5.85 * 2.73Polish−1.13 6.79 * 2.59Vocoded1.08 9.90 * 11.16 * French4 monthsNative2.50 17.84 * 8.33 * Polish1.52 17.25 * 7.19 * Vocoded−0.18 13.98 * 23.23 * 8 monthsNative−1.29 10.68 * 6.58 * Polish2.21 13.51 * 9.44 * Vocoded−0.33 11.38 * 26.94 * Japanese4 monthsNative0.66 19.86 * 4.82 * Polish−1.25 17.40 * 7.12 * Vocoded1.09 10.79 * 6.48 * 8 monthsNative0.03 16.44 * 0.39Polish−1.07 12.27 * 20.55 * Vocoded0.85 8.73 * 4.99 * Note: Asterisks (*) represent significant z‐scores (z‐scores > 3.1; p < 0.001, one‐tailed, i.e., signal > noise). z‐scores computed to determine whether significant EEG responses were elicited at the frequencies of interest (2.5, 5, and 10 Hz) for each stimulus type for each language and age group. Note: Asterisks (*) represent significant z‐scores (z‐scores > 3.1; p < 0.001, one‐tailed, i.e., signal > noise). Statistical Analyses Two Linear Mixed Effects (LME) models were constructed to address our two research questions. Model 1 assessed infants’ neural responses to native speech stimuli (English, French, and Japanese stimuli, respectively) as a function of their native language group (English, French, Japanese), age (4, 8 months), and frequency band (5 Hz, 10 Hz). Model 2 assessed infants’ neural responses to Polish and vocoded Polish stimuli as a function of stimulus type (speech, non‐speech), infants’ native language group (French, English, Japanese), age (4, 8 months), and frequency band (5 Hz, 10 Hz). Response amplitude was the dependent variable, which was scaled to facilitate the interpretation of the results, and contrast coding was used for all categorical independent variables (group: English 1, French 0, Japanese ‐1; age: 4mos 1, 8mos ‐1; frequency band: 5 Hz 1, 10 Hz ‐1; stimuli [model 2 only]: speech 1, non‐speech ‐1). The two models included random intercepts per participant. Analyses were conducted in R using the lme4 package (Bates et al. 2015) to run the models, the lmerTest package (Kuznetsova et al. 2017) to compute p values, and the emmeans package (Lenth 2016) to conduct pairwise comparisons to investigate the sources of any significant interactions (Tukey correction was used to adjust p values for multiple comparisons). Results Model 1: Neural Responses to Native Speech (English, French, or Japanese Stimuli) Neural responses to native speech by infants in the three language groups are shown in Figure 3, collapsed over age given the lack of age effects. Model 1 (Tables 2 and S1 for full model output) yielded a main effect of frequency band establishing that response amplitude overall was higher at 5 Hz compared to 10 Hz (B = 1.08, SE = 0.104, df = 113, t = 10.451, p < 0.001). There was also a significant frequency band by group interaction. Yet, pairwise comparisons showed that in spite of this interaction, the difference in responses at 5 Hz and 10 Hz, with greater response at 5 Hz, was significant in each of the language groups (English, B = 0.774, SE = 0.193, df = 113, t = 4.016, p < 0.001; French, B = 0.922, SE = 0.171, df = 113, t = 5.403, p < 0.001; Japanese, B = 1.559, SE = 0.175, df = 113, t = 8.894, p < 0.001). Response amplitude to native speech stimuli in the English, French, and Japanese groups (collapsed across age groups) in the 5 and 10 Hz frequency bands. TABLE 2Results of the ANOVA on the Linear Mixed Effects Model 1 that assessed neural response amplitude to native speech stimuli only as a function of frequency band, infants’ language group, and age (Response amplitude ∼Frequency * Group * Age + (1|Participant)).Sum SqMean SqNum dfDen dfF p Frequency668.7268.721113109.2284<0.001Group2.9821.49121132.36990.098Age0.1060.10611130.16910.682Frequency × Group6.7943.39721135.39920.006Frequency × Age00111300.999Group × Age0.2710.13621130.21540.807Frequency × Group × A0.2720.13621130.21620.806 Results of the ANOVA on the Linear Mixed Effects Model 1 that assessed neural response amplitude to native speech stimuli only as a function of frequency band, infants’ language group, and age (Response amplitude ∼Frequency * Group * Age + (1|Participant)). Model 2: Neural Responses to Speech Versus Non‐Speech (Polish vs. Vocoded Polish Stimuli) Neural responses to speech and non‐speech stimuli in the three language groups are shown in Figure 4. Model 2 (Tables 3 and S2 for full model output) yielded a main effect of frequency band. As for the speech only model reported above, response amplitude was significantly higher at 5 Hz than 10 Hz (ß = 0.933, SE = 0.073, t = 12.837, p < 0.001). The frequency band by stimulus type and group by stimulus type interactions were also significant. Response amplitude to speech and non‐speech stimuli in the English, French, and Japanese groups (collapsing across age groups) in the 5 and 10 Hz frequency bands. TABLE 3Results of the ANOVA on the Linear Mixed Effects Model 2 that assessed neural response amplitude as a function of frequency band, stimulus type (speech vs. non‐speech), infants’ language group and age (Response amplitude ∼ Frequency * Group * Age * Stimulus type + (1|Participant)).Sum SqMean SqNum dfDen dfF p Frequency101.618101.6181339164.796<0.001Group1.6640.83221131.3490.264Age0.0340.03411130.0540.816Stimuli0.8210.82113391.3310.249Frequency × Group3.0351.51723392.4610.087Frequency × Age1.2061.20613391.9560.163Group × Age1.7840.89221131.4470.240Frequency × Stimuli5.3795.37913398.7230.003Group × Stimuli7.2923.64623395.9130.003Age × Stimuli0.1310.13113390.2120.645Frequency× Group × Age0.9590.47923390.7770.460Frequency × Group× Stimuli0.9370.46923390.7600.469Frequency × Age × Stimuli0.2940.29413390.4770.490Group × Age × Stimuli1.2170.60823390.9870.374Frequency × Group × Age × Stimuli1.4570.72923391.1820.308 Results of the ANOVA on the Linear Mixed Effects Model 2 that assessed neural response amplitude as a function of frequency band, stimulus type (speech vs. non‐speech), infants’ language group and age (Response amplitude ∼ Frequency * Group * Age * Stimulus type + (1|Participant)). To understand the sources of these interactions, we conducted follow‐up LME models separately for each frequency band (see Tables 4 and S3 for full output of these models). Age was not included as a factor in these follow‐up models given that it did not yield any effects in Model 2. The model for the 5 Hz frequency band yielded a main effect of stimulus type: response amplitude was significantly higher for speech than non‐speech stimuli, B = 0.284, SE = 0.126, df = 116, t = 2.257, p = 0.026. The model for the 10 Hz frequency band, also yielded a main effect of stimulus type, but it was qualified by a stimulus type by group interaction. Pairwise comparisons showed that response amplitude was higher for speech than non‐speech stimuli in the Japanese group, B = 0.307, SE = 0.077, df = 116, t = 3.987, p < 0.001, though, unexpectedly, the reverse was found in the English, B = −0.257, SE = 0.084, df = 116, t = −3.044, p = 0.003, and French groups, B = −0.445, SE = 0.074, df = 116, t = −5.993, p < 0.001. TABLE 4Results of the ANOVAs on the follow‐up Linear Mixed Effects Models that assessed neural response amplitude as a function of stimulus type (speech vs. non‐speech) and infants’ language group separately for 5 Hz (Model 2A) and 10 Hz (Model 2B) (Response amplitude ∼ Group * Stimulus type + (1|Participant)).Model 2A: 5 HzSum SqMean SqNum dfDen dfF p Group3.1491.57521161.6880.189Stimuli4.7504.75011165.0930.026Group × Stimuli2.1411.07121161.1480.321Model 2B: 10HzSum SqMean SqNum dfDen dfF p Group0.3400.17021161.4010.251Stimuli1.0211.02111168.4250.004Group × Stimuli6.3633.181211626.245<0.001 Results of the ANOVAs on the follow‐up Linear Mixed Effects Models that assessed neural response amplitude as a function of stimulus type (speech vs. non‐speech) and infants’ language group separately for 5 Hz (Model 2A) and 10 Hz (Model 2B) (Response amplitude ∼ Group * Stimulus type + (1|Participant)). Discussion English‐, French‐, and Japanese‐learning 4‐ and 8‐month‐old infants were presented with sequences of 200‐ms‐long CVN (consonant‐vowel‐nasal) syllables, such that 2.5 Hz corresponded to the frequency of occurrence of bisyllables, 5 Hz to syllables, and 10 Hz to morae. First, our results establish in all three language groups and at both ages: (i) strong cortical tracking at both 5 and 10 Hz, but inconsistent tracking at 2.5 Hz (only in English‐learning 4‐month‐olds); (ii) better cortical tracking for 5 Hz than 10 Hz; and (iii) differences in cortical tracking for speech and non‐speech (vocoded) stimuli. In the following, we discuss implications of these findings for our understanding of the language‐general and language‐specific determinants of early cortical tracking of speech. The present results first provide clear evidence of statistically significant cortical tracking at different frequency bands simultaneously. We already know that the adult brain can track speech signals at different frequency bands simultaneously (Bröhl and Kayser 2021; Lizarazu et al. 2019; Peter et al. 2022; Rimmele et al. 2023; Teoh et al. 2019), so here we add to the growing evidence that this is also the case for infant brains (Attaheri et al. 2022a, 2022b; Ortiz Barajas et al. 2021,2023). What are the determinants of this cortical tracking in infancy? Adult results using the same stimuli (Peter et al. 2022) found both bottom‐up and top‐down factors. Bottom‐up effects were attested by the fact that more tracking was found at 5 than 10 than 2.5 Hz in French‐, English‐, and Japanese‐speaking adults, which mirrored the acoustic properties of the stimuli, in which there were higher acoustic peaks at 5 than 10 Hz, and no visible peak at 2.5 Hz. These acoustic properties likely resulted from the structure of the stimuli used, in which the regular syllabic structure (only CVN, using only CV and N morae, and no other types of morae) and the lack of prosodic marking of the foot level (all bisyllables being acoustically equal, corresponding to spondees). The present infant results also establish stronger tracking at 5 Hz than 10 Hz than 2.5 Hz, which is thus also in line with bottom‐up determinants of cortical tracking. Regarding top‐down effects, the previous study on adults found crosslinguistic modulation of cortical tracking, with a stronger syllabic advantage (5 Hz > 10 Hz) in the French‐speaking adults than in the English‐ and Japanese‐speaking adults (Peter et al. 2022). This modulation was interpreted as evidence that cortical tracking was also in part modulated by the rhythmic class of the language of adult listeners. However, no such language‐specific modulation was found in either the 4‐ or the 8‐month‐old infants in the present study. Although both age groups tracked rhythmic information in speech at both 5 Hz and 10 Hz, there were no age differences here nor were there any indications of language‐specific prosodic processing in either of these frequency bands. This does not match our predictions. Indeed, based on the behavioral literature showing language‐general rhythmic processing at 4 months, and an emergence of language‐specific rhythmic processing from 6 to 9 months onward (Jusczyk et al. 1993, 1999; Höhle et al. 2009; Skoruppa et al. 2009), we had predicted no cross‐linguistic modulation of cortical tracking at 4 months, but language‐specific modulation at 8 months, with a stronger syllabic advantage for the French‐learning infants given the adult findings of Peter et al. (2022). The present cortical tracking results do not converge with behavioral data, that is, there is no emergence of cross‐linguistic modulation at the cortical tracking level by 8 months. This also suggests that, given the adult results (Peter et al. 2022), changes in rhythmic processing at the neural level must occur later than 8 months. Despite these discrepancies, our findings align with current theoretical accounts of cortical tracking of speech in early language acquisition, which emphasize the critical role of tracking low‐frequency acoustic information (occurring at the delta and theta frequency bands) in the signal for successful encoding and segmentation of continuous speech (refer to the Temporal Sampling, Goswami 2015,2022). Thus, during the first months of life, infants’ brains track language‐general acoustic correlates to speech rhythm, which as a result of neural maturation and accumulated exposure to their native language (Peña et al. 2010; Menn et al. 2023), lead to the development of sensitivities to language‐specific rhythmic properties at the lexical level and acquisition of native phonology. Notably, the predominant frequency rate in our stimuli was 5 Hz, proposed to be the default rhythmic unit of adult‐directed speech across languages (Ding et al. 2017). Young infants, however, are primarily exposed to infant‐directed speech, which is characterized by predominant modulations at the delta band (∼2 Hz) and stronger delta‐theta phase coupling than adult‐directed speech, and this is the case for stress‐ (Leong et al. 2017), syllable‐ (Perez‐Navarro et al. 2022), and mora‐timed languages (Daikoku and Goswami 2025). Therefore, infants’ natural input likely provides bottom‐up cues to support the development of tracking at around 2 Hz. Therefore, it is possible that this first attempt to contrast stress‐, syllable‐, and mora‐based rhythms, using a duration manipulation (concatenations of 200 ms CVN syllables, made up of 2100 ms morae, potentially combining into 400 ms bisyllables), did not promote language‐specific processing, but rather a global perception of rhythm in which the 5 Hz rhythm, predominant across languages of the world and in our stimuli, was the one to which the brain entrained most strongly here. To further probe cortical tracking at the foot and mora levels, more natural stimuli might be used, in which the feet would be acoustically marked (e.g., by also manipulating loudness), or more mora types (e.g., long vowels or geminate consonants) would be used. Those more natural stimuli might increase language‐specific responses at the foot and mora levels in, respectively, English‐ and Japanese‐learning 8‐month‐olds (although a disadvantage of such stimuli would be that they would be less appropriate to test infants in the other languages). Finding such language‐specific effects could provide some light on our unexpected finding of cortical tracking at 2.5 Hz of the speech stimuli (English and Polish) in English‐learning infants at 4, but not 8 months, which might signal that while the 4‐month‐olds perceived bisyllables, the 8‐month‐olds did not as a result of having acquired that (trochaic) bisyllabes in English are marked by specific acoustic cues not present in our stimuli. Still related to our use of monotone sequences (no variability in intensity, duration and F0 of the syllables), it might be that some cortical tracking effects and modulations were partly masked, in line with data showing that infants better track stimuli that are more intonated (Kalashnikova et al. 2018; Menn, Michel, et al. 2022). Lastly, note that the age of 8 months roughly corresponds to the age at which language‐specific rhythmic perception begins to emerge, so it could be that some language specificity in our task (the more pronounced 5 Hz advantage in the French‐learning infants) could be found if (slightly) older infants were tested. Regarding the question of whether cortical tracking of speech differs from cortical tracking of non‐speech stimuli, it was found that adults in all three language groups showed no differences in speech versus non‐speech responses at 5 Hz, but better tracking for speech than for non‐speech at 10 Hz (Peter et al. 2022). The present infant results revealed a different pattern: an overall speech over non‐speech advantage at 5 Hz for the three language groups, and a speech versus non‐speech advantage at 10 Hz only for the Japanese‐learning infants (with a reversed effect at 10 Hz for the French‐ and the English‐learning infants). One possible interpretation of this pattern of results is that the better tracking of speech stimuli in infancy at the privileged frequency of 5 Hz is related to infants’ preference for speech over non‐speech stimuli found in behavioral studies (Vouloumanos and Werker 2007). Hence, increased cortical tracking at 5 Hz in infancy would signal easier processing of speech, as a favored kind of stimulus. This effect would presumably disappear at some point in development, with the brain beginning to process both speech and non‐speech similarly at 5 Hz. The results at 10 Hz—a speech advantage only by Japanese‐learning infants—are more challenging to interpret. Since 10 Hz corresponds to the mora‐based rhythm of Japanese, it is possible that experience with moraic rhythm in these infants facilitated their processing of speech at the moraic level, which would attest to some language‐specificity in tracking. However, we should be careful about such an interpretation, given that this effect is not strong enough to lead to better 10 Hz than 5 Hz tracking in these Japanese‐learning infants, that this effect was not found in adults (Peter et al. 2022), and that no language specificity was found when comparing 5 and 10 Hz processing across the three language groups, and might thus be a spurious effect given the sample size. An alternative explanation is that the speech versus non‐speech effects found in infants but not in adults, and the differential performance of the Japanese group compared to the English and French groups, were due to factors unrelated to rhythmic processing. For instance, attentional factors impact the efficiency of cortical tracking, and young infants, in particular, have been shown to track more efficiently stimuli that are more attention‐grabbing (e.g., infant‐directed speech, Kalashnikova et al. 2018; Menn, Michel, et al. 2022) or more familiar, such as their mother's voice (Jessen et al. 2025). This could be also the case for the speech versus non‐speech stimuli. In conclusion, this study is the first to evaluate cortical tracking in infants from both a crosslinguistic and a developmental perspective, testing infants at 4 and 8 months from three language backgrounds that exemplify the stress‐, syllable‐ and mora‐based rhythms. Our results establish that infant brains can track the speech signal at different frequencies simultaneously (5 and 10 Hz), with stronger tracking at 5 Hz, that corresponds to the syllabic level, the predominant rhythm across languages (Ding et al. 2017), and was acoustically more salient in our stimuli. Given that we found little evidence of developmental changes in cortical tracking between 4 and 8 months (except for the effect of tracking at 2.5 Hz at 4 but not 8 months in English‐learning infants for speech stimuli), and no evidence of crosslinguistic differences (except the Japanese speech versus non‐speech advantage at 10 Hz, and the English speech tracking at 2.5 Hz at 4 months), these results support the conclusion that there is bottom‐up, language‐general cortical tracking in early infancy at least until 8 months. Given previous findings of language‐specificity in tracking in adults (Peter et al. 2022), such differences should emerge some time before adulthood, presumably some time in infancy, after 8 months. This study also raises the possibility that the preferential tracking at 5 Hz may be present before 4 months, possibly at birth. In this case, studies with newborn infants and infants born prematurely would allow for the investigation of the possible influence of prosodic information available in utero during the last trimester of gestation on these effects. The results of this study, therefore, provide the basis for examining the full gamut of the development of early processing of linguistic prosody. Author Contributions Varghese Peter: conceptualization, investigation, writing – original draft, methodology, validation, visualization, software, formal analysis, data curation. Sandrien van Ommen: investigation, writing – review and editing, methodology, software, data curation, conceptualization. Denis Burnham: conceptualization, funding acquisition, writing – original draft, methodology, project administration, supervision, resources. Marina Kalashnikova: conceptualization, investigation, writing – original draft, methodology, validation, visualization, formal analysis, data curation. Thierry Nazzi: conceptualization, funding acquisition, writing – original draft, methodology, project administration, supervision, resources. Reiko Mazuka: conceptualization, funding acquisition, methodology, writing – review and editing, project administration, resources, supervision. Funding In Australia, this work was supported by a Transdisciplinary and Innovation Grant from Australian Research Council Centre of Excellence in Dynamics of Language. In France, it was funded by an ANR‐DFG grant (ANR‐16‐FRAL‐0007) and funds from LABEX EFL (ANR‐10‐LABX‐0083) to Thierry Nazzi, and a DIM Cerveau et Pensées grant (2013 MOBIBRAIN). In Japan, it was funded by JSPS grant‐in‐aid for Scientific Research S(16H06319) and for Specially Promoted Research (20H05617), MEXT grant for Innovative Areas #4903 (17H06382) to Reiko Mazuka. Ethics Statement The ethics committee for Human Research and Western Sydney University (Approval number: H9660), the ethics committee of CERES, France (No. 2011‐14), and the ethics committee for RIKEN (Wako 3 24‐11(20) and Wako 3 30‐2(1)∼(5)) approved the experimental methods used in the study. All experiments were performed in accordance with relevant guidelines and regulations. Informed consent was obtained from all the participants. Conflicts of Interest The authors declare no competing interests. Supporting information Supporting information: desc70271‐supp‐0001‐SuppMat.docx Acknowledgment Open access publication funding provided by COUPERIN CY26. Peter, V. , Kalashnikova M., Ommen S. van, Mazuka R., Burnham D., and Nazzi T.. 2026. “Cortical Tracking of Speech Across Rhythm Classes in English‐, French‐, and Japanese‐Learning 4‐ and 8‐Month‐Old Infants.” Developmental Science 29, no. 5: e70271. 10.1111/desc.70271 Varghese Peter and Marina Kalashnikova shared first authorship. Denis Burnham and Thierry Nazzi shared last authorship. Endnotes This differs from Peter et al. (2022), in which all adults were presented with the stimulus sequences spoken by each of the three voices – English, French and Japanese. Here, infants were only presented with the stimulus sequences spoken in their native language to reduce the duration of the experiment and accommodate their more limited attention span. Data Availability Statement Data cannot be uploaded to an open repository since we do not have ethical approval to do so, and we did not collect participants' consent for making data openly available. However, we will be able to share anonymized data upon request by researchers who have obtained the necessary ethical permissions to analyze these data. References Associated Data Supplementary Materials Supporting information: desc70271‐supp‐0001‐SuppMat.docx Data Availability Statement Data cannot be uploaded to an open repository since we do not have ethical approval to do so, and we did not collect participants' consent for making data openly available. However, we will be able to share anonymized data upon request by researchers who have obtained the necessary ethical permissions to analyze these data.