Predictive Neural Signals during Natural Mandarin Speech Comprehension

Predictive Neural Signals during Natural Mandarin Speech Comprehension | bioRxiv /* */ /* */ <!-- <!-- /*! * yepnope1.5.4 * (c) WTFPL, GPLv2 */ (function(a,b,c){function d(a){return"[object Function]"==o.call(a)}function e(a){return"string"==typeof a}function f(){}function g(a){return!a||"loaded"==a||"complete"==a||"uninitialized"==a}function h(){var a=p.shift();q=1,a?a.t?m(function(){("c"==a.t?B.injectCss:B.injectJs)(a.s,0,a.a,a.x,a.e,1)},0):(a(),h()):q=0}function i(a,c,d,e,f,i,j){function k(b){if(!o&&g(l.readyState)&&(u.r=o=1,!q&&h(),l.onload=l.onreadystatechange=null,b)){"img"!=a&&m(function(){t.removeChild(l)},50);for(var d in y[c])y[c].hasOwnProperty(d)&&y[c][d].onload()}}var j=j||B.errorTimeout,l=b.createElement(a),o=0,r=0,u={t:d,s:c,e:f,a:i,x:j};1===y[c]&&(r=1,y[c]=[]),"object"==a?l.data=c:(l.src=c,l.type=a),l.width=l.height="0",l.onerror=l.onload=l.onreadystatechange=function(){k.call(this,r)},p.splice(e,0,u),"img"!=a&&(r||2===y[c]?(t.insertBefore(l,s?null:n),m(k,j)):y[c].push(l))}function j(a,b,c,d,f){return q=0,b=b||"j",e(a)?i("c"==b?v:u,a,b,this.i++,c,d,f):(p.splice(this.i++,0,a),1==p.length&&h()),this}function k(){var a=B;return a.loader={load:j,i:0},a}var l=b.documentElement,m=a.setTimeout,n=b.getElementsByTagName("script")[0],o={}.toString,p=[],q=0,r="MozAppearance"in l.style,s=r&&!!b.createRange().compareNode,t=s?l:n.parentNode,l=a.opera&&"[object Opera]"==o.call(a.opera),l=!!b.attachEvent&&!l,u=r?"object":l?"script":"img",v=l?"script":u,w=Array.isArray||function(a){return"[object Array]"==o.call(a)},x=[],y={},z={timeout:function(a,b){return b.length&&(a.timeout=b[0]),a}},A,B;B=function(a){function b(a){var a=a.split("!"),b=x.length,c=a.pop(),d=a.length,c={url:c,origUrl:c,prefixes:a},e,f,g;for(f=0;f<d;f++)g=a[f].split("="),(e=z[g.shift()])&&(c=e(c,g));for(f=0;f<b;f++)c=x[f](c);return c}function g(a,e,f,g,h){var i=b(a),j=i.autoCallback;i.url.split(".").pop().split("?").shift(),i.bypass||(e&&(e=d(e)?e:e[a]||e[g]||e[a.split("/").pop().split("?")[0]]),i.instead?i.instead(a,e,f,g,h):(y[i.url]?i.noexec=!0:y[i.url]=1,f.load(i.url,i.forceCSS||!i.forceJS&&"css"==i.url.split(".").pop().split("?").shift()?"c":c,i.noexec,i.attrs,i.timeout),(d(e)||d(j))&&f.load(function(){k(),e&&e(i.origUrl,h,g),j&&j(i.origUrl,h,g),y[i.url]=2})))}function h(a,b){function c(a,c){if(a){if(e(a))c||(j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}),g(a,j,b,0,h);else if(Object(a)===a)for(n in m=function(){var b=0,c;for(c in a)a.hasOwnProperty(c)&&b++;return b}(),a)a.hasOwnProperty(n)&&(!c&&!--m&&(d(j)?j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}:j[n]=function(a){return function(){var b=[].slice.call(arguments);a&&a.apply(this,b),l()}}(k[n])),g(a[n],j,b,n,h))}else!c&&l()}var h=!!a.test,i=a.load||a.both,j=a.callback||f,k=j,l=a.complete||f,m,n;c(h?a.yep:a.nope,!!i),i&&c(i)}var i,j,l=this.yepnope.loader;if(e(a))g(a,0,l,0);else if(w(a))for(i=0;i (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0];var j=d.createElement(s);var dl=l!='dataLayer'?'&l='+l:'';j.src='//www.googletagmanager.com/gtm.js?id='+i+dl;j.type='text/javascript';j.async=true;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-M677548'); Skip to main content Home About Submit ALERTS / RSS Search for this keyword Advanced Search New Results Predictive Neural Signals during Natural Mandarin Speech Comprehension Qifei Wang , Eva Berlot , Judit Fazekas , Jakub Szewczyk , View ORCID Profile Floris P de Lange doi: https://doi.org/10.1101/2025.11.23.690006 Qifei Wang 1 Donders Institute for Brain, Cognition and Behaviour, Radboud University Nijmegen , Nijmegen, the Netherlands Find this author on Google Scholar Find this author on PubMed Search for this author on this site For correspondence: qifei.wang{at}donders.ru.nl Eva Berlot 1 Donders Institute for Brain, Cognition and Behaviour, Radboud University Nijmegen , Nijmegen, the Netherlands Find this author on Google Scholar Find this author on PubMed Search for this author on this site Judit Fazekas 1 Donders Institute for Brain, Cognition and Behaviour, Radboud University Nijmegen , Nijmegen, the Netherlands Find this author on Google Scholar Find this author on PubMed Search for this author on this site Jakub Szewczyk 2 Institute of Psychology, Jagiellonian University , Kraków, Poland Find this author on Google Scholar Find this author on PubMed Search for this author on this site Floris P de Lange 1 Donders Institute for Brain, Cognition and Behaviour, Radboud University Nijmegen , Nijmegen, the Netherlands 3 Centre for Artificial Intelligence and Neuroscience, University of Bonn , Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Floris P de Lange Abstract Full Text Info/History Metrics Preview PDF Abstract Language comprehension requires transforming continuous speech into hierarchical linguistic representations, from phonemes to words to sentences. Prediction has been proposed as a key mechanism supporting this process across languages. Here, we extend this framework to Mandarin Chinese, a tonal language with a distinct phonological structure. Using magnetoencephalography (MEG), we recorded neural responses from native speakers listening to an audiobook, and applied linear regression modeling to examine how linguistic features modulate brain activity. We found that listeners segmented continuous speech into multiple linguistic units, and that surprisal modulated neural responses at several representational levels, though not at the finest phonemic level. Tonal information was not independently predicted, but played a critical role when integrated with its segmental component. Finally, entropy reduction—an information-theoretic measure that quantifies how much a word reduces uncertainty about future words in a sentence—elicited a later and temporally distinct neural response from surprisal, indicating independent contributions of these two information-theoretic measures. Together, our findings suggest that predictive mechanisms in language comprehension are universal in their computational principles, but implemented in ways shaped by language-specific structural features, and that local surprisal resolution precedes the global updating of sentence interpretation. Introduction When listening to speech, the brain performs a remarkable feat: it transforms a continuous acoustic stream of sound into a structured, meaningful message. This process begins with the segmentation of the incoming signal into discrete linguistic units–such as phonemes, syllables, words, and phrases ( Brent, 1999 ; Mattys et al., 2005 ; Sanders & Neville, 2000 ). The brain tracks the time courses of these abstract structures in parallel ( Ding et al., 2016 ; Giraud & Poeppel, 2012 ), and interprets them incrementally as speech unfolds ( Marslen-Wilson, 1973 ). However, because natural speech is rapid ( Crystal & House, 1990 ) and often degraded by noise ( McDermott, 2009 ), it is essential for the brain to actively predict upcoming content, to ensure that signal can be disambiguated in real time and to maintain pace with the speaker without delay ( Clark, 2013 ; Pickering & Garrod, 2013 ). Substantial evidence supports this predictive mechanism during speech and across multiple linguistic levels ( Heilbron et al., 2022 ; Kuperberg & Jaeger, 2016; Ryskin & Nieuwland, 2023 ; Tezcan et al., 2023 ). For instance, while listening to an audiobook (e.g., the boy went out to fly a…), the brain actively predicts not only the next word (’kite’; Weissbart et al., 2020 ; Willems et al., 2016 ), but also the next phoneme (’/k/’; Brodbeck et al., 2022 ; DeLong et al., 2005 ; Donhauser & Baillet, 2020 ) based on prior context. Yet, despite being considered a universal feature of language processing, predictive mechanisms have been investigated only in a handful of languages. How predictive processing generalizes across typologically distinct languages therefore remains unknown. To address this gap, the current study uses Mandarin Chinese—the world’s most widely spoken tonal language with a unique phonological and orthographical system—to investigate several key questions about predictive processing during naturalistic speech comprehension. First, we aimed to determine the level(s) of granularity at which linguistic segmentation and prediction operate. While Mandarin speech segmentation likely occurs concurrently at multiple levels, as has been suggested for English ( Giraud & Poeppel, 2012 ), we asked whether predictive processing also functions across all levels of representations, or whether it is reducible to a single, primary level of representation, for example the one carrying the greatest semantic relevance. The linguistic structure of Mandarin naturally defines four levels of representation: phonemes, the smallest contrastive units that change a word’s meaning ( Chomsky & Halle, 1968 ); sub-syllabic units, which can be composed of multiple phonemes, and serve as crucial instructional tool in Chinese literacy education; characters, the basic units of the writing system that typically correspond to both a syllable and a morpheme; and words, composed of one or more characters that convey semantically complete meaning. This hierarchical structure offers a clear framework to examine how segmentation and prediction unfold across different linguistic levels. In addition, we also examined the role of lexical tone in Mandarin prediction. Tone is a fundamental phonological feature in Mandarin, as it distinguishes meaning among syllables with identical segmental structures (i.e., the same initial and final sequence). For example, /ta1/ means 她 (“she”), whereas /ta3/ means 塔 (“tower”). While phonological prediction has been extensively documented in non-tonal languages ( DeLong et al., 2005 ; Ito et al., 2020 ; Näätänen et al., 1997 ; Van Petten et al., 1999 ), evidence for tonal prediction in Mandarin remains scarce. Behavioral studies employing object-naming paradigms have not found evidence for tone-specific prediction ( Xu et al., 2024 ; Zhao et al., 2024 ). More recent work, nevertheless, has shown that tonal information can be anticipated in highly constrained contexts, with decodable neural signal reflecting tone prediction (Liu, Chen, et al., 2025; Liu, Zhang, et al., 2025). Similarly, violations of expected tone reliably elicit robust neural responses ( Brown-Schmidt & Canseco-Gonzalez, 2004 ; Huang et al., 2018 ; Zhao et al., 2011 ). However, these studies used rather constrained and artificial settings, and it thus remains unknown whether listeners also actively form tonal predictions during natural speech comprehension. To examine how continuous speech is segmented into discrete linguistic units, we tested the onset effect at different levels of granularity, since prior work demonstrated that onset-evoked responses are reliable markers of speech segmentation ( Sanders et al., 2002 ; Sanders & Neville, 2003 ). To assess predictive processing across these linguistic dimensions—including tone—we leveraged surprisal , a widely used computational proxy of linguistic prediction based on information theory ( Levy, 2008 ; Shain et al., 2024 ). Surprisal quantifies how unexpected an input (e.g., a word) is, given its preceding context. With the advent of large language models like GPT-2, it can now be estimated at a scale not achievable with traditional cloze probability tests ( Taylor, 1953 ). Words with high surprisal typically impose greater processing demand and elicit stronger neural responses, particularly around 400 ms after word onset ( Kutas & Federmeier, 2000 ; Kutas & Hillyard, 1980 ; Van Petten & Luka, 2012 ). Lastly, we ask a more general question: does the brain’s response to a word only reflects the word’s surprisal, or does it also reflect how much a word shapes the overall interpretation of the whole sentence ( Hale, 2006 )? For this, we computed entropy reduction (ER), another information-theoretic metric that quantifies how much the encountered word decreases uncertainty about the possible sentence continuation ( Hale, 2003 , 2006 ). While both metrics predict how much neural processing each word may require, surprisal captures the immediate information provided by the word itself, whereas ER reflects the word’s broader impact on updating listener’s interpretation of the discourse ( Hale, 2016 ). Despite prior evidence that surprisal and ER independently predict reading times ( Frank, 2013 ), the specific neural consequences of ER are not well understood ( Frank et al., 2015 ). This may partly be due to the limitations of earlier studies, which relied on less sophisticated language models and presented words visually in isolation—conditions that may bias toward immediate next-word prediction while restricting longer-range inferences. To address all these questions, we analyzed magnetoencephalography (MEG) recordings from native Mandarin speakers listening to a Chinese audiobook. We quantified linguistic predictions with a Chinese GPT-2 model in a fine-grained, context-sensitive manner and used a time-resolved multivariate temporal response function (mTRF) analysis ( Crosse et al., 2016 ; Di Liberto et al., 2015) to capture the temporal dynamics of the neural responses to multiple features in naturalistic/continuous signal. To preface the results, we found that although speech segmentation and predictive processing were both visible at multiple linguistic levels, there were marked differences between them. Whereas speech segmentation was most strongly visible at the level of phonemes, surprisal was most prominent at more abstract, higher linguistic levels of analysis. Tonal prediction did not emerge independently but contributed critically when integrated with its segmental information. Finally, we found that ER was represented in the brain independently and with a distinct temporal profile from surprisal, demonstrating that these two metrics capture complementary dimensions of predictive language processing. Together, these findings suggest that while predictive mechanisms are universal in their computational principles, their implementation is tailored to language-specific structural features, and that local prediction resolution precedes the global updating of sentence interpretation. Results Thirty-four native Mandarin speakers listened to 51 minutes of a Mandarin audiobook while their brain activity was recorded using magnetoencephalography (MEG). Participants were instructed to listen attentively while fixating on a central cross. To ensure attention, the story was split into 40 segments, each followed by a short recognition task in which participants judged which of two short clips had appeared in the preceding segment (see Method). Performance was high (mean accuracy = 95.3%), confirming sustained engagement. To examine whether and how neural responses encode different linguistic features (e.g., character surprisal), we employed multivariate temporal response functions (mTRFs) modeling. This technique estimates the linear relationship between fluctuations in speech features and continuous MEG signals ( Fig. 1 ; see Methods). Because speech unfolds rapidly (e.g., characters occur on average every 200 ms), neural responses to adjacent events overlap in time. The mTRF framework addresses this by estimating the contribution of each linguistic regressor across a range of time lags, thereby disentangling overlapping responses. To assess whether specific linguistic features were neurally represented, we hierarchically added regressors—based on their individual contribution to model fit—and assessed whether each subsequent addition yielded a significant improvement, as determined using cross-validation approach (i.e. correspondence of prediction of left-out data compared to real MEG signal; Fig.1 ). We focused our analyses primarily on temporal sensors, given prior evidence implicating temporal regions as critical for language processing ( Gagnepain et al., 2012 ; Willems et al., 2016 ). In addition, we employed a Bayesian one-sample t test to evaluate the hemispheric lateralization of the observed effects ( Rouder et al., 2009 ; see Methods). Download figure Open in new tab Figure 1. Experimental paradigm and analysis pipeline. Participants listened to a Chinese audiobook while their brain activity was recorded using magnetoencephalography (MEG). The linguistic structure of Mandarin is illustrated: words comprise one or more characters; each character maps to a monosyllable, which is composed of sub-syllabic units—an optional initial (consonant) and a final (vowel(s), sometimes with a nasal consonant) that carries a lexical tone (indicated by numbers 1-4). Some sub-syllabic units can be further segmented into phonemes. The audiobook transcript was analyzed using a Chinese GPT-2 model to quantify contextual surprisal for each character. Temporal response function (TRF) modelling was used to estimate the linear relationship between linguistic features and MEG signal. As illustrated for character-level surprisal, the regressor was time-shifted (steps of 1/120 s, spanning 100 ms before to 1000 ms after character onset) to construct time-lagged regressors. TRF model estimates the beta weights at each time lag, capturing the unique contribution of this surprisal feature to the MEG signal. Model performance was evaluated via cross-validation, calculating the Pearsons’ correlation ( r ) between the predicted (black) and observed (grey) neural responses on held-out data. Speech is segmented at multiple levels of granularity We first investigated the neural segmentation of continuous speech across four linguistic levels: phonemic, sub-syllabic, character, and word levels. In Mandarin, the character serves as the basic unit of the writing system and typically corresponds to a single syllable and a morpheme (the basic unit of meaning in the spoken language). Each syllable is composed of an initial consonant (optional), a final , and a tone ( Cheng, 2011 ; e.g., 她 corresponds to syllable /t-a1/; Fig. 1 ). The final contains one or more vowels, sometimes followed by a nasal consonant. Critically, these sub-syllabic units are the primary perceptual units taught in Chinese reading, whereas phonemes, though the smallest meaning-distinguishing units, are not explicitly emphasized in literacy education. It is noteworthy though, that our participants were fluent English speakers, a language where phonemic awareness is central ( Mesgarani et al., 2014 ). Moreover, while individual characters carry meaning, they are often semantically ambiguous in isolation, and most words are formed by combining two or more characters ( Pan et al., 2021 ; Fig. 1 ). This linguistic structure provides a clear hierarchy of four grain sizes for our examination: phoneme, sub-syllabic, character and word. To control for potential confounds arising from low-level acoustic properties, we first constructed a baseline model incorporating acoustic features—such as the broadband envelope, acoustic edges, pitch, and spectrogram—as impulse regressors (see Methods). We then added regressors of interest: onsets at the phonemic, sub-syllabic, character, and word levels, defined also as impulses. Given the inherent temporal overlap and correlation among these onset regressors, we employed a hierarchical procedure, sequentially adding the regressor that yielded the greatest improvement in model fit at each step. This approach allowed us to determine whether each successive level of linguistic structure provided a significant and unique explanation of the neural response beyond those already included in the model. The results showed that adding the phoneme onset predictor was the most important in capturing neural variance ( t (33) = 17.3, p < 1 × 10 -40 , d = 2.96 ; Fig. 2A ). The second biggest contribution was the sub-syllabic onset level predictor ( t (33) = 12.3, p = 3.77 × 10 -14 , d = 2.11; Fig. 2A ). Next, word onset still provided a further improvement in model fit ( t (33) = 9.55, p = 2.55 × 10 -11 , d = 1.64; Fig. 2A ). In contrast, character onset did not contribute unique variance beyond other levels ( t (33) = 1.32, p = 0.097, d = 0.23; Fig. 2A ). These results suggest that the brain concurrently tracks boundaries across multiple levels of linguistic granularity, from fine-grained phonemic structure to semantically defined words. Download figure Open in new tab Figure 2. Neural representations of speech onsets at different linguistic levels. (A) Model performance when adding regressors with different onset types. Bars represent prediction accuracy for models that included different regressors of onsets. The baseline model contains acoustics, including envelope, spectrogram, acoustic edge, and pitch. Each ‘ + ’ model included one onset regressor on top of the previous model. For example, the + sub-syllabic model contains acoustics + phoneme onset + sub-syllabic onset . The order of the added regressors are determined based on their individual contribution to model performance. Dots with connecting lines represent individual participants (averaged over all temporal sensors). (B) Temporal profiles of the global field power (GFP) of the beta coefficients from the full onset model (i.e., the + character model), plotted separately for each onset type. (C) Topographic maps of beta coefficients at two peak time windows, shown separately for phoneme, sub-syllabic, character, and word onsets. We next investigated the temporal dynamics and spatial profile of these neural responses. Inspecting the global field power of the estimated coefficients from the full TRF onset model across temporal sensors (see Methods), we identified two prominent deflections following each onset type. The first peak, occurring around 100 ms, showed consistent right-hemisphere dominance across all linguistic levels, as quantified by the hemispheric lateralization index (LI, see Methods; phoneme: mean LI = −0.11, BF 10 = 96.11; sub-syllabic: −0.13, BF 10 = 587.19; character: −0.11, BF 10 = 152.39; word: −0.10, BF 10 = 9.63; Fig. 2B-C ). The second deflection, peaking around 200 ms, only exhibited right-hemisphere lateralization for character onsets (phoneme: −0.06, BF 10 = 0.65; sub-syllabic: −0.06, BF 10 = 0.95; character: −0.09, BF 10 = 4.42; word: −0.06, BF 10 = 0.60; Fig. 2B-C ) . Moreover, the relative amplitude of the second deflection, compared to the first, increased systematically from the phoneme to the word level (F (3, 132) = 10.7, p = 2.43 × 10 -6 , η 2 = 0.20; Fig. 2B ). Post-hoc tests revealed that this amplitude difference was significantly larger for character and word onsets than for phoneme and sub-syllabic onsets. Language prediction occurs at multiple levels of granularity Our finding that Chinese speech is segmented at multiple linguistic levels raises an important question: does language prediction also occur simultaneously across these levels? To address this, we examined prediction at the phoneme, sub-syllabic, character, and word levels. We quantified linguistic prediction using surprisal, derived from Chinese GPT-2 ( Radford et al., 2019 ; Zhao et al., 2019 , 2023 ), an autoregressive language model trained to predict the next character given prior context. GPT-2 assigns a probability to each predicted token (character) based on its contextual likelihood, from which we computed surprisal (−log(p)) for every character in the audiobook transcript ( Fig. 1 ). To determine the optimal context length, we estimated multiple TRF models using character-level surprisal generated under different context windows (from a single character, up to the model’s maximum of 1024 characters). We found that longer context window better captured the brain’s response to surprisal (linear fit: β = 3.68 × 10⁻⁴, t (33) = 12.5, p = 2.14 × 10 -14 , Fig. 3D ), echoing prior behavioral findings that human predictions of upcoming words align more closely with GPT-2’s predictions as context length increases ( Goldstein et al., 2022 ). Accordingly, we used the maximal context window of 1024 for all further analyses. From these character-level estimates, we derived surprisal values for other linguistic levels. Word surprisal was calculated from the joint probability of its constituent characters; sub-syllabic surprisal (initial and final) was computed from summing the probabilities of predicted characters sharing the same initial, or the same initial-final sequence, conditioned on the initial; phonemic surprisal was computed from summing the probabilities of predicted characters sharing the same phoneme sequence, conditioned on the preceding phoneme (see Methods). Download figure Open in new tab Figure 3. Surprisal effects at different levels of granularity. (A) Model performance across different surprisal regressor conditions. The baseline (onset) model included acoustic features, all onset regressors, and character frequency. Each ‘ + ’ model included one surprisal regressor on top of the previous model. The order of the added regressors are determined based on their individual contribution to model performance. (B) Temporal profiles of the global field power (GFP) of the beta coefficients from the full surprisal model (i.e., + word model), plotted separately for each surprisal. (C) Topographic maps of beta coefficients at the time windows of 200-400 ms, shown separately for phoneme, sub-syllabic, character, and word surprisal. (D) Model performance as a function of context window size used in Chinese GPT-2 to estimate character surprisal, the models included acoustics + all onsets + character frequency + character surprisal . Following the onset-modeling approach, we incrementally added surprisal regressors according to their unique contribution to model fit, while controlling for acoustic features, all onset regressors, and character frequency, i.e., how frequent the character is in a large text corpus ( Cai & Brysbaert, 2010 ). Frequency was included to ensure that any effects we find are not driven by infrequent characters, but instead truly reflect context-driven surprisal effects ( Gillis et al., 2021 ). Our results showed that at all levels except for phonemic, i.e. the finest granularity level, surprisal was a significant predictor ( Fig. 3A ) . Sub-syllabic surprisal explained the most variance ( t (33) = 11.9, p = 8.10 × 10 -14 , d = 2.05). Adding character surprisal further improved model fit ( t (33) = 8.98, p = 1.11 × 10 -10 , d = 1.54), and word surprisal also uniquely added to model predictability ( t (33) = 2.31, p = 0.014, d = 0.40). In contrast, phoneme surprisal did not improve (but rather worsened) model fit ( t (33) = −3.08, p = 0.99, d = 0.52), suggesting that predictive processing in Mandarin Chinese operates concurrently at multiple levels of granularity but not necessarily at the finest phonemic level. The corresponding TRFs revealed robust N400-like responses, particularly for sub-syllabic (mean latency = 273 ms) and character (mean latency = 306 ms) surprisal. The peak latency for sub-syllabic surprisal was significantly earlier than for character surprisal (jackknife-based latency t test: t (33) = 2.9, p = 0.0065, d = 0.50; Fig. 3B ). Spatial analyses provided moderate evidence against hemispheric asymmetry across all levels (mean LI = 0.03, 0.02, −0.02, −0.02 for phoneme, sub-syllabic, character and word, respectively; all BF 10 = 0.2; Fig. 3C ), indicating bilateral engagement in predictive processing in Mandarin. Thus, a key dissociation emerged between onset and surprisal effects. While phoneme onset significantly modulated neural responses, phoneme surprisal did not contribute uniquely to brain activity. A further dissociation was observed at the higher levels: word onset was prioritized over character onset during segmentation ( s Fig. 1A ) , whereas the reverse pattern held for surprisal—character surprisal exerted a stronger effect than word surprisal ( s Fig. 1B ). Together, these dissociations suggest that linguistic prediction does not necessarily operate at the finest detectable level (phoneme), and that the brain weighs linguistic units differently for speech segmentation and prediction, optimizing processing for distinct functional goals. Tone and rhyme predictions in Mandarin Chinese Next, we examined whether listeners generate predictions about lexical tones during speech comprehension. As lexical tones in Mandarin are primarily realized over the rhyme portion of the syllable—the vowel(s) and any following nasal consonant(s)—we evaluated predictive processing at both the tone and rhyme levels. Tone surprisal and rhyme surprisal were derived from character-level surprisal by considering the conditional probabilities of GPT-predicted characters sharing the same initial and either the same tone −log( p (tone | initial, context) ) or the same rhyme −log( p (rhyme | initial, context) ), respectively (see Methods; Fig. 4A ). Additionally, to assess whether tone and rhyme jointly carry predictive information beyond their individual contributions, we calculated an integrated tone-rhyme surprisal −log( p (tone, rhyme | initial, context) ) that captures their joint contributions ( Fig. 4A ). Download figure Open in new tab Figure 4. Tone and rhyme predictions. (A) Illustration of probability estimation for different phonological components. GPT-2 generated predictions for upcoming characters with varying probabilities; the target character is marked in red. Each character corresponds to a syllable comprising an initial (green), rhyme (muted purple), and tone (dark magenta). Initial probability was computed by summing the probabilities of all characters sharing the same initial (e.g., all ‘ch’ characters). Rhyme and tone probabilities were calculated by summing probabilities of characters that share either both the same initial and rhyme (e.g. ch–i) or the same initial and tone (e.g. ch–1), then normalized by the initial probability. Integrated probability was computed with characters that match the target in initial, rhyme, and tone (e.g. ch–i1), normalized by the probability of the initial. (B) Model performance across different surprisal regressor conditions. The tone + rhyme model contained tone and rhyme surprisal, whereas + tone x rhyme model included tone surprisal, rhyme surprisal and integrated tone and rhyme surprisal. (C) Temporal profiles of the global field power (GFP) of the beta coefficients from the full model (i.e., the + tone x rhyme model), plotted separately for each surprisal type. (D) Topographic maps of beta coefficients in different time windows, shown separately for tone, rhyme and integrated surprisal. We fit TRF models that incorporated tone surprisal, rhyme surprisal, or both, alongside a baseline with acoustic and onset regressors. The results showed that adding rhyme surprisal to the model with tone surprisal significantly improved the prediction of neural responses ( t (33) = 10.9, p = 8.79 × 10 -13 , d = 1.87; Fig. 4B ). Conversely, adding tone surprisal to the rhyme-only model did not yield a significant improvement ( t (33) = 0.93, p = 0.18, d = 0.16; Fig. 4B ), suggesting no independent tonal prediction. Crucially, when the integrated tone-rhyme surprisal was added to a model containing both tone and rhyme surprisal additively, model fit improved substantially ( t (33) = 7.29, p = 1.15 × 10 -8 , d = 1.25; Fig. 4B ). This indicates that the joint contribution of tone and rhyme carries unique predictive information not captured by their separate, additive effects. Temporal analyses revealed an early neural deflection at approximately 100 ms for both tone and rhyme surprisal ( Fig. 4C ), accompanied by right-lateralized topographies (tone: mean LI = −0.06, BF 10 = 1.37; rhyme: −0.12, BF 10 = 1.1× 10 4 ; Fig. 4D ). In contrast, the integrated tone-rhyme surprisal induced a markedly stronger neural response that was more bilaterally distributed (mean LI = 0.01, BF 10 = 0.2; Fig. 4C-D ). Dissociable effects of surprisal and entropy reduction So far, we have examined how surprisal at different linguistic levels modulates neural responses during Mandarin speech comprehension. Next, we investigated whether entropy reduction (ER) modulates neural activity separately from surprisal. While surprisal reflects the local unexpectedness of the current character, ER quantifies how much that character reduces uncertainty about upcoming characters, thereby capturing its global role in constraining the space of likely future interpretations. Following Frank (2013) , we operationalized ER using a 3-character lookahead window, estimating how much the current character reduces uncertainty about the next three characters ( Fig. 5A ; see Methods). The 3-character window was chosen to balance computation efficiency with explanatory power, and ER computed within this range has been shown to robustly modulate reading times independent of surprisal ( Frank, 2013 ). We focused our analyses on the character level, given that characters in Mandarin are the primary carriers of meaning, making them highly informative units for updating predictions about upcoming content. We constructed separate TRF models with regressors for surprisal and ER, as well as an additive model including both, to isolate their unique neural signatures. Download figure Open in new tab Figure 5. Dissociable effects of surprisal and entropy reduction. (A) Illustration of entropy reduction (ER) computation. For each character, we computed the uncertainty associated with the probable continuations of the sentence (up to 3-characters) before the character was presented (blue histogram) and after its presentation (up to 2 characters – red histogram). More specifically, the top-k predicted characters generated by GPT-2 were used (blue circles; k was set to cover 0.9 cumulative probability, capped at 40 for computational efficiency). First, we computed the entropy before the actual character (red circle) was presented by tracing along each of the predicted characters the potential upcoming context recursively to generate possible upcoming 3-character sequences. The probability of each sequence was calculated as the product of its three component character probabilities, forming a probability distribution (blue distribution). The entropy of this distribution reflected uncertainty about the upcoming continuation before the character was presented. Next, we computed the entropy after the actual character appeared, i.e. how the presentation of the character updated the possible sentence continuations. We followed the same procedure, but this time calculating possible 2-character sequences, forming a second distribution (red distribution). The difference between the two entropy values was taken as the entropy reduction (ER) for the target character. (B) Model performance across different regressor conditions. (C) Temporal profiles of the global field power (GFP) of the beta coefficients from the surprisal + ER model, plotted separately for each regressor type. (D) Topographic maps of beta coefficients for the two time windows, shown separately for ER and surprisal. Both measures contributed independently to explaining neural responses. Adding surprisal to a model containing ER significantly improved model fit ( t (33) = 14.0, p = 9.41 × 10 -16 , d = 1.80; Fig. 5B ). Conversely, adding ER to a model containing surprisal also yielded a significant gain ( t (33) = 3.9, p = 2.09 × 10 -4 , d = 0.71; Fig. 5B ). This mutual independence confirms that surprisal and ER represent distinct types of predictive information during speech comprehension. Temporal analyses revealed a clear dissociation: the neural response to surprisal peaked at 326 ms after character onset, whereas the response to ER peaked significantly later, at 446 ms (jackknife-based latency t test: t (33) = 7.2, p = 3.30 × 10 -8 , d = 1.2; Fig. 5C ). Spatially, both surprisal and ER effects were primarily bilateral at their peaks, respectively (surprisal: mean LI = −0.004, BF 10 = 0.2 for 200-400 ms; ER: −0.02, BF 10 = 0.22 for 400-600 ms; Fig. 5D ). Discussion In this study, we examined how language processing modulates brain activity during naturalistic Mandarin speech comprehension by applying linear modelling to MEG signals. Our findings reveal a fundamental dissociation between speech segmentation and predictive processing: while the brain tracks linguistic boundaries and generates predictions at multiple sizes of granularity, prediction (as measured by surprisal) operates mostly at higher levels of linguistic analysis, despite robust neural sensitivity to the onsets of lowest-level, phonemic units. This suggests that predictive processing might be tailored to language-specific structure. Furthermore, tonal information alone does not drive prediction, but strengthens it when integrated with phonological structure, highlighting the interdependence of segmental (rhyme) and suprasegmental (tone) cues. Lastly, entropy reduction (ER) is independently represented and temporally dissociable from surprisal, suggesting prediction and uncertainty updating constitute distinct yet complementary computation in the predictive brain. Our first finding is that the listening brain is sensitive to linguistic onsets spanning from phoneme to word, indicating parallel segmentation of continuous speech into units of different sizes. This supports theoretical frameworks proposing that listeners use a hierarchy of cues to segment speech ( Brent, 1999 ; Mattys et al., 2005 ). Neurophysiological evidence from English confirms that the brain concurrently tracks phoneme and word boundaries ( Karunathilake et al., 2023 , 2025 ), even within noisy conditions ( Brodbeck et al., 2018 ). Our data extend these findings to Mandarin, demonstrating parallel tracking from phonemes to words, and crucially, revealing Chinese native’s sensitivity to boundaries of language-specific sub-syllabic units (initials and finals). This effect may arise from the acoustic and articulatory properties of Mandarin finals: the transitions between phonemes are temporally compressed and coarticulated, such that multiple vowels or nasals are perceived as a cohesive perceptual chunk. This temporal compactness may foster perceptual binding at the sub-syllabic level. At the morpho-lexical level, we observed preferential sensitivity to word onsets over character onsets, suggesting that the brain prioritizes meaningfulness over granularity—segmenting speech into semantically complete words more than often-ambiguous single characters, thereby promoting efficient comprehension ( Li et al., 2009 ). Such multi-level segmentation provides complementary information, facilitating robust and rapid processing in the face of noise and variability ( Ding & Simon, 2014 ; Gross et al., 2013 ; Keitel et al., 2018 ; Saberi & Perrott, 1999 ). In addition, we observed two deflections following all onset types. The first (∼100 ms) resembles the N100, a well-established neural marker of word onset effect in both artificial and naturalistic language ( Brodbeck et al., 2018 ; Sanders et al., 2002 ; Sanders & Neville, 2003 ; Tezcan et al., 2023 ). The second deflection (∼200 ms) exhibited a graded increase in amplitude relative to the first from phoneme and sub-syllabic to character and word levels. This graded pattern likely reflects the integration of semantic information: as characters and words carry morphemic or lexical meaning, binding these features into the evolving discourse model may demand greater computational resources ( Broderick et al., 2018 , 2019 ; Huang et al., 2014 ; Van Den Brink et al., 2001 ). Next, our results demonstrate that the brain is responsive to surprisal at multiple levels of granularity, with each level capturing unique predictive information derived from the prior context. This supports theoretical proposals that predictive processing operates across variable timescales and levels of representational abstraction ( Altmann & Mirković, 2009 ), and replicates empirical findings in English ( Gillis et al., 2021 ; Heilbron et al., 2022 ). Critically, we found that predictive signals were predominantly driven by sub-lexical surprisal. This aligns with the fundamental incrementality of language comprehension: the brain does not wait for a complete lexical unit to unfold but continuously refines expectations at each sub-lexical step, generating rapid and robust “micro-predictions” ( Brodbeck et al., 2018 ; Donhauser & Baillet, 2020 ). This multi-level predictive architecture allows the brain to efficiently resolve ambiguities and redundancies present at different stages of processing, thereby guiding the interpretation of continuous speech with remarkable speed and robustness ( Christiansen & Chater, 2016 ). A key finding is the absence of a unique surprisal effect at the phoneme level, despite clear neural sensitivity to phoneme boundaries. This dissociation suggests that although the brain can segment speech into universal, fine-grained phonemic units, predictive processing relies on language-specific, functionally salient units. In Mandarin, sub-syllabic units (initials and finals) are the foundational units of literacy education and pronunciation, whereas phonemes are not explicitly taught. Some neural studies indicate that phonemes constitute fundamental units of phonological encoding in Mandarin ( Qu et al., 2012 , 2020 ), but behavioral work suggested phonemes play a subordinate role in the preparation of word production ( Chen et al., 2016 ; O’Seaghdha et al., 2010 ). Critically, these studies exploited only the first phoneme, ignoring the final. Evidence from Cantonese, a closely related Chinese dialect, shows that shared finals can drive priming ( Wong & Chen, 2008 , 2009 ), underscoring the functional relevance of the sub-syllabic unit. Moreover, sub-syllabic awareness is a critical predictor of early reading acquisition in Chinese ( Shu et al., 2008 ; Siok & Fletcher, 2001 ). From a distributional perspective, the higher transitional probabilities within finals, relative to those between initial-final boundaries, and the high frequency of co-occurrence among phonemes within finals, likely reinforce sub-syllabic units as the stable processing units ( Hay & Baayen, 2005 ). Together, these findings suggest that while phonemes provide a universal articulatory scaffold, the sub-syllabic chunk constitutes the behaviorally and cognitively relevant unit for generating predictions during Mandarin speech comprehension. We observed a further dissociation at the morpho-lexical level: word-level segmentation was prioritized (word onsets over character onsets), yet character-level prediction was more robust (character surprisal over word surprisal). This indicates that the brain optimizes different processes for complementary goals. Segmentation prioritizes disambiguating the input into semantically complete words to ensure processing efficiency, whereas predictive processing relies more heavily on incremental updating and thus on the most recent meaningful morphemic unit (the character) to enable flexible and rapid integration. This balance between stability and flexibility enables listeners to maintain coherent meaning representations while continuously updating expectations in response to new sensory evidence. Furthermore, we found that tonal prediction does not operate independently in natural Mandarin speech comprehension, whereas prediction of the rhyme (i.e., the non-initial element of a syllable) uniquely explains the neural signal. This aligns with previous findings from eye-tracking, which used the visual world paradigm and failed to find anticipatory fixations toward tonal distractors, observing only marginal effects for rhyme, suggesting that rhyme—but not tone—is pre-activated predictively ( Xu et al., 2024 ; Zhao et al., 2024 ). In contrast, ERP studies showed that explicit tonal violations can elicit N400 effects ( Brown-Schmidt & Canseco-Gonzalez, 2004 ; Hu et al., 2012 ; Zhao et al., 2011 ). However, these paradigms often use semantically incongruent or lexically implausible stimuli, which may reflect semantic violation effects rather than genuine tonal prediction. By quantifying surprisal across all contextually possible continuations weighted on their possibilities, our approach provides a more ecologically valid measure of tonal prediction. Our finding also aligns with evidence that rhyme information is a stronger constraint than tone in lexical access ( Sereno & Lee, 2015 ; Wiener & Turnbull, 2016 ; Yang & Chen, 2022 ), especially in predictive context ( Shen et al., 2021 ). A recent EEG study reporting decodable tonal predictions (Liu, Zhang, et al., 2025) appears to contradict our null result. We posit that this discrepancy stems from fundamental methodological differences. Their task explicitly required participants to predict one of two highly constrained tones, strongly encouraging strategic tonal prediction. In contrast, during naturalistic listening to an audiobook, the candidate space is vastly larger and unconstrained, making independent tonal prediction less essential and its neural signature more subtle, thus potentially harder to detect in aggregate brain responses. Most importantly, we found that the integrated tone-rhyme surprisal provided more predictive power than the sum of the its parts. This synergistic effect demonstrates that while tone may not be predicted in isolation, it carries information that is indispensable when combined with its segmental counterpart. This is consistent with behavioral findings of increased anticipatory fixations to distractors sharing both tone and rhyme with a target ( Li et al., 2022 ; Xu et al., 2024 ; Zhao et al., 2024 ). Together, during natural comprehension, listeners do not generate independent predictions about tone; they integrate tonal and segmental information to guide efficient interpretation of speech. Lastly, we demonstrated that ER and surprisal modulate neural responses in distinct ways. Although both metrics measure the information conveyed by a word and assume that greater information value corresponds to greater processing effort, they capture different aspects of information processing ( Hale, 2016 ). Surprisal measures the unexpectedness of each incoming word, whereas ER reflects the reduction of uncertainty of a few words ahead, linking comprehension to overall sentence meaning disambiguation effort ( Chen & Hale, 2021 ; Hale, 2003 ). Prior studies have demonstrated that ER influences reading times independently of surprisal ( Frank, 2013 ; Linzen & Jaeger, 2016 ), and engages shared ( Nelson et al., 2017 ) but also distinct neural networks during naturalistic listening ( Song et al., 2024 ). Computational linguistic model also predicts the dissociation between ER and surprisal in incremental comprehension ( Venhuizen et al., 2019 ). Although ERP studies using reading paradigms have struggled to disentangle ER from surprisal ( Frank et al., 2015 ), ER has been hypothesized to correspond either to the N400, like surprisal, or to a later components such as the P600 ( Frank et al., 2015 ; Graben et al., 2000 ; Hale, 2016 ). By leveraging large language model for precise information-theoretic estimation and an ecologically valid naturalistic paradigm, our findings provide neural evidence that ER and surprisal are temporally dissociable during Mandarin speech comprehension. Specifically, ER elicited a later neural deflection distinct from the N400-like effect associated with surprisal, suggesting that listeners first resolve local contextual conflicts and subsequently integrate information to update their internal predictive model of the unfolding linguistic context. This pattern indicates that listeners not only anticipate incoming content but also actively infer the meaning of the sentence by allocating greater processing resources to words that are more informative about the sentence—reflecting brain’s intrinsic drive for information gain. While our results provide compelling evidence for the distinct representation of ER and surprisal, future work employing more sensitive ER indices of sentence meaning updating will be crucial for further validating this dissociation. From a broader perspective, these findings may offer insight for improving computational language models. Modern large language models (LLMs) have achieved impressive linguistic competence, yet still fall short on tasks requiring deeper functional reasoning ( Mahowald et al., 2024 ). They are trained primarily to minimize cross-entropy loss ( Goodfellow et al., 2016 ), which is mathematically equivalent to minimizing expected surprisal. Given the parallel between human predictive comprehension and probabilistic next-word prediction in LLMs ( Goldstein et al., 2022 ), incorporating an additional ER-inspired objective could, in principle, help models better capture the dynamics of human-like uncertainty management. Such an approach might encourage models not only predict accurately but also to optimize information gain over time, reflecting how the human brain integrate prediction with uncertainty reduction during comprehension. Although speculative, the idea aligns with emerging computational evidence that post-training with the objective of minimizing token-level entropy alone enhances LLMs’ reasoning performance ( Agarwal et al., 2025 ) and LLMs predicting sentences leads to more human-like representations ( Yu et al., 2024 ), suggesting a promising avenue for future interdisciplinary research. In summary, our findings illustrate the universality of predictive processing while demonstrating its adaptation to language-specific structures. Prediction does not necessarily engage every phonological sublevel equally, but operates more efficiently at those levels that maximize interpretive efficiency. Once predictions are resolved, listeners integrate the new input to refine their inferences of the ongoing communication. Materials and methods Data availability All data and code used for stimulus presentation and analysis will be freely available on the Donders Repository after a FAIR review protocol. Participants Thirty-four native speakers of Mandarin Chinese were recruited (female = 21, aged from 21-36 years, mean = 28 years, SD = 3.73 years). None reported any neurological, developmental, or language impairments. All participants were right-handed, with normal hearing and normal or corrected-to-normal vision. They provided written informed consent for their data to be used for research purposes and received monetary compensation for their participation. The study was approved by the local ethics committee (CMO 2014/288; CMO Arnhem-Nijmegen, the Netherlands) and conducted in accordance with the Declaration of Helsinki. The sample size was determined based on a priori power analysis ( Faul et al., 2009 ), which indicated a sample of 34 would achieve 0.8 power to detect a medium effect size (Cohen’s d = 0.5) at ɑ = 0.05 (two-tailed paired t test). Stimuli and procedure All audio stimuli were presented using Psychtoolbox and played through a loudspeaker positioned directly in front of the participant at a clear and comfortable listening volume. Participants sat comfortably while their brain activity was continuously recorded using magnetoencephalography (MEG). The listening material consisted of the opening chapters of the audiobook Bailuyuan , split into 40 segments ranging from 56 to 91 seconds in duration (totaling 50.7 minutes and 9,945 characters). All recordings were presented during a single experimental session. Segments were presented in chronological order to preserve contextual continuity. Following each segment, participants performed a brief recognition task. They heard two short audio clips (each ∼3 seconds), one from the segment they had just heard and the other from a later, unused portion of the audiobook. The two clips were presented in random order, and participants indicated which one had appeared in the preceding segment using a button press. Four buttons were used and responded to with the index and middle fingers of each hand: the left hand indicating the first clip and the right hand the second. Confidence was reported by the exact finger: middle fingers indicated high confidence and index fingers low confidence. Feedback on the correctness of participants’ choice was provided after each response. The task was intentionally simple for native Mandarin speakers and served primarily to maintain attention during listening. After each trial, participants could take a break and start the next segment at their own pace. Stimuli annotation To obtain precise onset times for each character and sub-syllabic unit, the manually transcribed text was aligned with the audio using the Montreal Forced Aligner (MFA; McAuliffe et al., 2017 ). This alignment provided onset and offset timestamps for each character and each phoneme. All automatically generated onsets and offsets were then visually inspected and manually adjusted in Praat ( Boersma & Weenink, 2011 ) to ensure temporal accuracy. Word boundaries were determined using HanLP ( He & Choi, 2021 ) toolkit for word segmentation, which was then visually inspected for correction, resulting in 5880 words. Word onsets were defined as the start time of the word’s first character. Each character was converted to its phonological form using the pinyin Python package ( Huang et al., 2025 ). The sub-syllabic structure of each character was then parsed: the initial (where present) was treated as the first sub-syllabic unit, and the final with tone as the second. The onset times for all sub-syllabic units were assigned based on phoneme-level alignment obtained from MFA. MEG acquisition and preprocessing Brain responses were recorded with a whole-head 275-channel axial gradiometer MEG system (CTF) inside a magnetically shielded room at the Donders Centre for Cognitive Neuroimaging in Nijmegen, the Netherlands. MEG signals were digitized at a sampling rate of 1200 Hz with an online 300-Hz low-pass filter applied during acquisition. Participants’ head positions were continuously monitored using localization coils placed at the nasion and the left and right pre-auricular points. If head movement exceeded 5 mm from the initial position, participants were asked to readjust it during breaks. In addition, participants’ left eyes were tracked using an SR Research Eyelink 1000 eye tracker, and behavioral responses were collected using MEG-compatible button boxes. MEG data were preprocessed and analyzed using the Fieldtrip toolbox in Matlab (version 2023b; Oostenveld et al., 2011 ). Continuous recordings were segmented into 40 trials based on the timestamps of the 40 audio segments. A notch filter (Butterworth IIR) was applied at 49-51, 99-101, and 149-151 Hz to remove potential line noise artifacts. Next, data was band-pass filtered between 0.1 - 20 Hz using a bidirectional finite impulse response filter to focus on slow, evoked responses ( Heilbron et al., 2022 ). Data were then downsampled to 120 Hz and cleaned for artifacts such as muscle contractions and SQUID jumps using a semi-automated approach. Specifically, trials were divided into 500 ms snippets with 200 ms overlap between them. Snippets with unusually high variance were visually identified and removed. The cleaned snippets were then combined to reconstruct the 40 trials and demeaned within trials to remove any DC drift. Finally, independent component analysis was performed to remove remaining artifacts caused by eye movements, heartbeat, or other non-neural sources of noise. Predictor regressors Onset regressors. Onset regressors were constructed as impulses with a value of 1 at the onset of each language unit–phoneme, sub-syllabic unit, character and word–and 0 elsewhere. To verify that onset effects were not spurious, we included a fake word regressor constructed by randomly sampling phoneme onsets; this control regressor contained the same number of sample points as the true word onsets. Each regressor spanned the entire length of the neural signal. Regressors of prediction. Linguistic prediction was estimated using Chinese GPT-2 ( Zhao et al., 2019 ), a large pretrained language model that performs comparably to humans in next-word prediction tasks. The model’s basic token is the character. The text of the audiobook was passed through GPT-2 to obtain conditional probability of each character based on its preceding context. To determine the optimal context window for probability estimation, we generated predictions using windows of 1, 2, 3, 8, 16, 32, 64, 128, 256, 512 and 1024 tokens. The 1024-character window provided the best model fit and was therefore used for all subsequent analyses. These character-level probabilities were used to derive probabilities for words, sub-syllabic units, phonemes, tones, and rhymes. Specifically, word probabilities were computed as the joint probabilities of their component characters based on the word boundaries. Sub-syllabic probabilities were estimated by summing the probabilities of those predicted characters sharing the same initial (first) or both the initial and final (second) components, adjusting for the sequential order (i.e., the probability of the second sub-syllabic unit was conditional on the first). Similarly, phonemic probabilities were estimated by summing the probabilities of those predicted characters sharing the same phoneme up to that point, and normalized by the probability of the preceding phoneme. Tone and rhyme probabilities were derived also in a similar way, marginalizing over predicted characters that shared the relevant phonological feature and normalizing appropriately. At all levels, only the top k predictions whose cumulative probability reached at least 0.9 were taken into consideration, with k set to a minimum of 40 to ensure sufficient information even in low-entropy contexts ( Heilbron et al., 2022 ). Surprisal for each unit was computed as the negative log probability (-log p). Entropy reduction (ER) was computed following Frank (2013) , using a 3-character lookahead window to approximate uncertainty about future input. For each target character, likely continuations were recursively generated to form 3-character sequences, and sequence probabilities were calculated to estimate pre-target entropy. After updating the context with the target character, entropy was recomputed for the next 2-character sequences, and difference yielded the ER value. If a predicted continuation terminated in a full stop, the recursion was ended early and the available predictions were used for computing that sequence probability. Although a sentence-level estimate would be preferable, exhaustive computation of all possible sentence continuations is computationally intractable. The 3-character window thus follows Frank’s approach, which has been shown to robustly predict reading behavior. To balance coverage ad efficiency, the number of predicted characters for each estimate was constrained to a maximum of 40 and a minimum of 10. Regressors of no interest. To ensure that the observed effects were not confounded by low-level acoustic or other factors, we included several control variables known to drive brain responses: broadband envelope, spectrogram, acoustic edges, pitch, and character frequency. The original audio was first downsampled to 120 Hz after applying a zero-phase anti-aliasing filter. The broadband amplitude envelope was calculated as the absolute value of the sum of the raw speech signal and its Hilbert transform. The spectrogram was derived by filtering the speech stimuli into 8 frequency bands between 250 Hz and 8 kHz according to Greenwoord’s equation ( Greenwood, 1961 ), then computing the amplitude envelope within each band as before. Including both the envelope and spectrogram enables the model to account for spectral and temporal variations in the acoustic signal (Di Liberto et al., 2015). To capture transient changes in amplitude (acoustic edges; Daube et al., 2019 ), we included the standard deviation of each phoneme’s amplitude from onset to offset, based on boundaries defined by the MFA. Pitch was calculated for each phoneme using Praat via the parselmouth interface, as prosodic variation has been shown to modulate prediction-related neural responses ( Zhang et al., 2021 ). All acoustic control variables were modeled as impulses aligned with phoneme onsets. Finally, to control for character frequency, we included the unigram surprisal of each character, defined as -log p(character) based on its frequency of occurrence in SUBTLEX-CH ( Cai & Brysbaert, 2010 ). Regression analysis To analyze the mapping between the various linguistic features and the recorded MEG signal, we used multiple temporal response functions (mTRF) – a method widely adopted in naturalistic language research ( Crosse et al., 2016 ). MTRFs accounts for the substantial temporal overlap between brain reactions to adjacent events (e.g., phonemes). Conceptually, a TRF can be viewed as a linear filter that describes how the brain transforms a stimulus feature, S(t), into a continuous neural response R(t), i.e., R(t) = TRF*S(t), where * represents the convolution operator. All predictors were temporally expanded into time-lagged regressors to model the evoked MEG response. Each lag was assigned a weight, capturing the contribution of that predictor to the brain response at that particular latency. To stabilize estimates in the presence of multicollinearity (both between predictors and across their lags), we applied ridge regression, which penalizes large weights. This allows the model to isolate distinct, time-resolved contribution of each predictor, even when their neural effects overlap in time. In this way, mTRF provides a continuous estimate of neural coding, analogous to an event-related field (ERF/ERP), but with the crucial advantage of explicitly accounting for the temporal overlap and confounds that traditional ERF analyses struggle with in continuous naturalistic input. The time range of lagging was restricted from 100 pre-event onset to 1000 ms post-event onset (−100—1000 ms) at a step of 1/120 second for analysis, as no reliable response was observed outside this range. MEG signals and all predictors except the binary onset regressors were z-scored per trial prior to fitting. The MEG data for the mTRF analysis were in the axial gradiometer configuration, which preserved both positive and negative signal polarities. However, because the orientation and location of the underlying neural sources can vary across individuals, averaging the resulting beta coefficients directly across subjects may obscure true effects. To address this, the sensor-level beta coefficients were transformed into planar gradients before averaging. This transformation approximates the local neural response as maximal directly above its neural source, producing positive-only values that are more interpretable and comparable across participants ( Hämäläinen et al., 1993 ). Finally, to visualize the overall temporal profile across sensors, we computed the global field power using only the group-averaged beta coefficients from the temporal sensors. Model comparison Model performance was quantified by comparing cross-validated correlation coefficients (Pearson’s r) between predicted and observed MEG signals. We used a 10-fold leave-one-out cross-validation scheme, where trials were randomly partitioned once into 10 folds for each subject. In each fold, 36 trials were used for training and 4 for testing. Crucially, for each subject, the same train/test partitions were applied to all models, ensuring that performance differences reflected the predictors rather than differences in trial assignment. Predictive performance was averaged across folds to yield a robust estimate of model generalizability. Statistical testing All statistical tests, unless specified, were one-tailed with an alpha level of 0.05. For within-subject comparisons of model performance, we used paired t tests. To compare the amplitude difference between the two onset peaks across all linguistic levels, we calculated for each subject the amplitude difference of the peaks in the first (50-150 ms) and the second (150-300 ms) windows across the four onset type, and employed analysis of variance on them. Post-hoc comparisons used Tukey’s HSD test. To compare peak latencies between surprisal and entropy reduction, we employed a jackknife-based latency t test, which estimates individual peak latencies by iteratively leaving out one participant and recomputing group-level statistics ( Miller et al., 1998 ). To assess whether model performance for surprisal varies linearly with context window size (from 1 to 1024 characters), we fit a linear regression for each participant, with model performance as a function of context window length, and tested whether the resulting slopes differed from zero using a one-sample t test. To examine hemispheric lateralization, we computed laterality index (LI) for each subject, defined as LI = (L-R)/(L+R), where L and R represent the planar-converted beta coefficients in the left and right temporal sensors, respectively. Positive LI value indicates left-hemisphere lateralization, whereas negative LI value indicates right-hemisphere lateralization. We then tested whether the LI differed from zero using a Bayesian one-sample t test implemented in pingouin Python package ( Vallat, 2018 ). The Bayesian approach was used because it provides a continuous measure of evidence for or against the null hypothesis (no hemispheric lateralization), enabling more nuanced interpretation and avoiding dichotomous thresholding ( Wagenmakers et al., 2018 ) lateralization. By convention, BF 10 > 10 indicates strong evidence for lateralization, 3 < BF 10 < 10 indicates moderate evidence, 1/3 < BF 10 < 3 indicates anecdotal evidence, and BF 10 < 1/3 indicates moderate evidence for the absence of lateralization. Download figure Open in new tab Supplementary Figure 1. Comparison between models with single linguistic predictors added to different baseline models. (A) Adding word onsets to the baseline model significantly improved performance relative to adding character onsets, indicating that word boundaries were prioritized during speech segmentation. Baseline 1 included all acoustic regressors and two sub-lexical onset regressors (phonemic and sub-syllabic onsets). (B) Adding character surprisal to the baseline model significantly improved performance relative to adding word surprisal, indicating that characters were more influential than words in linguistic prediction. Baseline 2 included all acoustic and onset regressors, as well as the sub-syllabic surprisal regressor. All regressors of interest were added on top of their corresponding baseline models. Each dot represents data from an individual participant. Asterisks indicate statistical significance (p < 0.01, p < 0.001). References ↵ Agarwal , S. , Zhang , Z. , Yuan , L. , Han , J. , & Peng , H . ( 2025 ). The Unreasonable Effectiveness of Entropy Minimization in LLM Reasoning (No . arXiv : 2505 . 15134 ). arXiv. doi: 10.48550/arXiv.2505.15134 OpenUrl CrossRef ↵ Altmann , G. T. M. , & Mirković , J . ( 2009 ). Incrementality and Prediction in Human Sentence Processing . Cognitive Science , 33 ( 4 ), 583 – 609 . doi: 10.1111/j.1551-6709.2009.01022.x OpenUrl CrossRef PubMed ↵ Boersma , P. , & Weenink , D . ( 2011 ). Praat: Doing Phonetics by Computer . Ear and Hearing , 32 ( 2 ), 266 . doi: 10.1097/AUD.0b013e31821473f7 OpenUrl CrossRef ↵ Brent , M. R . ( 1999 ). Speech segmentation and word discovery: A computational perspective . Trends in Cognitive Sciences , 3 ( 8 ), 294 – 301 . doi: 10.1016/S1364-6613(99)01350-9 OpenUrl CrossRef PubMed Web of Science ↵ Brodbeck , C. , Bhattasali , S. , Cruz Heredia , A. A. , Resnik , P. , Simon , J. Z. , & Lau , E . ( 2022 ). Parallel processing in speech perception with local and global representations of linguistic context . eLife , 11 , e72056 . doi: 10.7554/eLife.72056 OpenUrl CrossRef PubMed ↵ Brodbeck , C. , Hong , L. E. , & Simon , J. Z . ( 2018 ). Rapid Transformation from Auditory to Linguistic Representations of Continuous Speech . Current Biology , 28 ( 24 ), 3976 – 3983 .e5. doi: 10.1016/j.cub.2018.10.042 OpenUrl CrossRef PubMed ↵ Broderick , M. P. , Anderson , A. J. , Di Liberto , G. M. , Crosse , M. J. , & Lalor , E. C. ( 2018 ). Electrophysiological Correlates of Semantic Dissimilarity Reflect the Comprehension of Natural, Narrative Speech . Current Biology , 28 ( 5 ), 803 – 809 .e3. doi: 10.1016/j.cub.2018.01.080 OpenUrl CrossRef PubMed ↵ Broderick , M. P. , Anderson , A. J. , & Lalor , E. C . ( 2019 ). Semantic Context Enhances the Early Auditory Encoding of Natural Speech . Journal of Neuroscience , 39 ( 38 ), 7564 – 7575 . doi: 10.1523/JNEUROSCI.0584-19.2019 OpenUrl Abstract / FREE Full Text ↵ Brown-Schmidt , S. , & Canseco-Gonzalez , E . ( 2004 ). Who Do You Love, Your Mother or Your Horse? An Event-Related Brain Potential Analysis of Tone Processing in Mandarin Chinese . Journal of Psycholinguistic Research , 33 ( 2 ), 103 – 135 . doi: 10.1023/B:JOPR.0000017223.98667.10 OpenUrl CrossRef PubMed Web of Science ↵ Cai , Q. , & Brysbaert , M . ( 2010 ). SUBTLEX-CH: Chinese Word and Character Frequencies Based on Film Subtitles . PLOS ONE , 5 ( 6 ), e10729 . doi: 10.1371/journal.pone.0010729 OpenUrl CrossRef PubMed ↵ Chen , J.-Y. , O’Séaghdha , P. G. , & Chen , T.-M . ( 2016 ). The primacy of abstract syllables in Chinese word production. Journal of Experimental Psychology: Learning , Memory, and Cognition , 42 ( 5 ), 825 – 836 . doi: 10.1037/a0039911 OpenUrl CrossRef ↵ Chen , Z. , & Hale , J. T . ( 2021 ). Quantifying Structural and Non-structural Expectations in Relative Clause Processing . Cognitive Science , 45 ( 1 ), e12927 . doi: 10.1111/cogs.12927 OpenUrl CrossRef ↵ Cheng , C.-C . ( 2011 ). A Synchronic Phonology of Mandarin Chinese . De Gruyter Mouton . doi: 10.1515/9783110866407 OpenUrl CrossRef ↵ Chomsky , N. , & Halle , M. ( 1968 ). The sound pattern of English . ↵ Christiansen , M. H. , & Chater , N . ( 2016 ). The Now-or-Never bottleneck: A fundamental constraint on language . Behavioral and Brain Sciences , 39 , e62 . doi: 10.1017/S0140525X1500031X OpenUrl CrossRef PubMed ↵ Clark , A . ( 2013 ). Whatever next? Predictive brains, situated agents, and the future of cognitive science . Behavioral and Brain Sciences , 36 ( 3 ), 181 – 204 . doi: 10.1017/S0140525X12000477 OpenUrl CrossRef PubMed ↵ Crosse , M. J. , Di Liberto , G. M. , Bednar , A. , & Lalor , E. C. ( 2016 ). The Multivariate Temporal Response Function (mTRF) Toolbox: A MATLAB Toolbox for Relating Neural Signals to Continuous Stimuli . Frontiers in Human Neuroscience , 10 . doi: 10.3389/fnhum.2016.00604 OpenUrl CrossRef PubMed ↵ Crystal , T. H. , & House , A. S . ( 1990 ). Articulation rate and the duration of syllables and stress groups in connected speech . The Journal of the Acoustical Society of America , 88 ( 1 ), 101 – 112 . doi: 10.1121/1.399955 OpenUrl CrossRef PubMed Web of Science ↵ Daube , C. , Ince , R. A. A. , & Gross , J . ( 2019 ). Simple Acoustic Features Can Explain Phoneme-Based Predictions of Cortical Responses to Speech . Current Biology , 29 ( 12 ), 1924 – 1937 .e9. doi: 10.1016/j.cub.2019.04.067 OpenUrl CrossRef PubMed ↵ DeLong , K. A. , Urbach , T. P. , & Kutas , M . ( 2005 ). Probabilistic word pre-activation during language comprehension inferred from electrical brain activity . Nature Neuroscience , 8 ( 8 ), 1117 – 1121 . doi: 10.1038/nn1504 OpenUrl CrossRef PubMed Web of Science Di Liberto , G. M. , O’Sullivan, J. A., & Lalor, E. C. ( 2015 ). Low-Frequency Cortical Entrainment to Speech Reflects Phoneme-Level Processing . Current Biology , 25 ( 19 ), 2457 – 2465 . doi: 10.1016/j.cub.2015.08.030 OpenUrl CrossRef PubMed ↵ Ding , N. , Melloni , L. , Zhang , H. , Tian , X. , & Poeppel , D . ( 2016 ). Cortical tracking of hierarchical linguistic structures in connected speech . Nature Neuroscience , 19 ( 1 ), 158 – 164 . doi: 10.1038/nn.4186 OpenUrl CrossRef PubMed ↵ Ding , N. , & Simon , J. Z . ( 2014 ). Cortical entrainment to continuous speech: Functional roles and interpretations . Frontiers in Human Neuroscience , 8 . doi: 10.3389/fnhum.2014.00311 OpenUrl CrossRef PubMed ↵ Donhauser , P. W. , & Baillet , S . ( 2020 ). Two Distinct Neural Timescales for Predictive Speech Processing . Neuron , 105 ( 2 ), 385 – 393 .e9. doi: 10.1016/j.neuron.2019.10.019 OpenUrl CrossRef PubMed ↵ Faul , F. , Erdfelder , E. , Buchner , A. , & Lang , A.-G . ( 2009 ). Statistical power analyses using G*Power 3.1: Tests for correlation and regression analyses . Behavior Research Methods , 41 ( 4 ), 1149 – 1160 . doi: 10.3758/BRM.41.4.1149 OpenUrl CrossRef PubMed Web of Science ↵ Frank , S. L . ( 2013 ). Uncertainty Reduction as a Measure of Cognitive Load in Sentence Comprehension . Topics in Cognitive Science , 5 ( 3 ), 475 – 494 . doi: 10.1111/tops.12025 OpenUrl CrossRef PubMed ↵ Frank , S. L. , Otten , L. J. , Galli , G. , & Vigliocco , G . ( 2015 ). The ERP response to the amount of information conveyed by words in sentences . Brain and Language , 140 , 1 – 11 . doi: 10.1016/j.bandl.2014.10.006 OpenUrl CrossRef PubMed ↵ Gagnepain , P. , Henson , R. N. , & Davis , M. H . ( 2012 ). Temporal Predictive Codes for Spoken Words in Auditory Cortex . Current Biology , 22 ( 7 ), 615 – 621 . doi: 10.1016/j.cub.2012.02.015 OpenUrl CrossRef PubMed ↵ Gillis , M. , Vanthornhout , J. , Simon , J. Z. , Francart , T. , & Brodbeck , C . ( 2021 ). Neural Markers of Speech Comprehension: Measuring EEG Tracking of Linguistic Speech Representations, Controlling the Speech Acoustics . The Journal of Neuroscience , 41 ( 50 ), 10316 – 10329 . doi: 10.1523/JNEUROSCI.0812-21.2021 OpenUrl Abstract / FREE Full Text ↵ Giraud , A.-L. , & Poeppel , D . ( 2012 ). Cortical oscillations and speech processing: Emerging computational principles and operations . Nature Neuroscience , 15 ( 4 ), 511 – 517 . doi: 10.1038/nn.3063 OpenUrl CrossRef PubMed ↵ Goldstein , A. , Zada , Z. , Buchnik , E. , Schain , M. , Price , A. , Aubrey , B. , Nastase , S. A. , Feder , A. , Emanuel , D. , Cohen , A. , Jansen , A. , Gazula , H. , Choe , G. , Rao , A. , Kim , C. , Casto , C. , Fanda , L. , Doyle , W. , Friedman , D. , … Hasson , U . ( 2022 ). Shared computational principles for language processing in humans and deep language models . Nature Neuroscience , 25 ( 3 ), 369 – 380 . doi: 10.1038/s41593-022-01026-4 OpenUrl CrossRef PubMed ↵ Goodfellow , I. , Bengio , Y. , Courville , A. , & Bengio , Y . ( 2016 ). Deep learning (Vol. 1, Issue 2) . MIT press Cambridge . ↵ Graben , P. B. , Saddy , J. D. , Schlesewsky , M. , & Kurths , J . ( 2000 ). Symbolic dynamics of event-related brain potentials . Physical Review E , 62 ( 4 ), 5518 – 5541 . doi: 10.1103/PhysRevE.62.5518 OpenUrl CrossRef PubMed ↵ Greenwood , D. D . ( 1961 ). Auditory masking and the critical band . Journal of the Acoustical Society of America , 33 , 484 – 502 . doi: 10.1121/1.1908699 OpenUrl CrossRef Web of Science ↵ Gross , J. , Hoogenboom , N. , Thut , G. , Schyns , P. , Panzeri , S. , Belin , P. , & Garrod , S . ( 2013 ). Speech Rhythms and Multiplexed Oscillatory Sensory Coding in the Human Brain . PLOS Biology , 11 ( 12 ), e1001752 . doi: 10.1371/journal.pbio.1001752 OpenUrl CrossRef PubMed ↵ Hale , J . ( 2003 ). The Information Conveyed by Words in Sentences . Journal of Psycholinguistic Research , 32 ( 2 ), 101 – 123 . doi: 10.1023/A:1022492123056 OpenUrl CrossRef PubMed ↵ Hale , J . ( 2006 ). Uncertainty About the Rest of the Sentence . Cognitive Science , 30 ( 4 ), 643 – 672 . doi: 10.1207/s15516709cog0000_64 OpenUrl CrossRef PubMed ↵ Hale , J . ( 2016 ). Information-theoretical Complexity Metrics . Language and Linguistics Compass , 10 ( 9 ), 397 – 412 . doi: 10.1111/lnc3.12196 OpenUrl CrossRef ↵ Hämäläinen , M. , Hari , R. , Ilmoniemi , R. J. , Knuutila , J. , & Lounasmaa , O. V . ( 1993 ). Magnetoencephalography—Theory, instrumentation, and applications to noninvasive studies of the working human brain . Reviews of Modern Physics , 65 ( 2 ), 413 – 497 . doi: 10.1103/RevModPhys.65.413 OpenUrl CrossRef Web of Science ↵ Hay , J. B. , & Baayen , R. H . ( 2005 ). Shifting paradigms: Gradient structure in morphology . Trends in Cognitive Sciences , 9 ( 7 ), 342 – 348 . doi: 10.1016/j.tics.2005.04.002 OpenUrl CrossRef PubMed Web of Science ↵ He , H. , & Choi , J. D. ( 2021 ). The Stem Cell Hypothesis: Dilemma behind Multi-Task Learning with Transformer Encoders . In M.-F. Moens , X. Huang , L. Specia , & S. W. Yih (Eds.), Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (pp. 5555 – 5577 ). Association for Computational Linguistics . doi: 10.18653/v1/2021.emnlp-main.451 OpenUrl CrossRef ↵ Heilbron , M. , Armeni , K. , Schoffelen , J.-M. , Hagoort , P. , & de Lange , F. P. ( 2022 ). A hierarchy of linguistic predictions during natural language comprehension . Proceedings of the National Academy of Sciences , 119 ( 32 ), e2201968119 . doi: 10.1073/pnas.2201968119 OpenUrl CrossRef PubMed ↵ Hu , J. , Gao , S. , Ma , W. , & Yao , D . ( 2012 ). Dissociation of tone and vowel processing in Mandarin idioms . Psychophysiology , 49 ( 9 ), 1179 – 1190 . doi: 10.1111/j.1469-8986.2012.01406.x OpenUrl CrossRef PubMed Web of Science ↵ Huang , H. , Kong , X. , MSK, I., bors-homu, Chen, W., Yang, Y., hanabi1224, Wang, D., bot, S., Gates, T., wolfgitpr, Lee, A., DavidBekkers, Timperi, J., Wang, M., Serfend, Xu, S., Badger, T. G., Yang, W., … wzy. ( 2025 ). mozillazg/python-pinyin: V0.54.0 [Computer software] . Zenodo . doi: 10.5281/zenodo.15108473 OpenUrl CrossRef ↵ Huang , X. , Liu , X. , Yang , J.-C. , Zhao , Q. , & Zhou , J . ( 2018 ). Tonal and vowel information processing in Chinese spoken word recognition: An event-related potential study . NeuroReport , 29 ( 5 ), 356 . doi: 10.1097/WNR.0000000000000973 OpenUrl CrossRef PubMed ↵ Huang , X. , Yang , J.-C. , Zhang , Q. , & Guo , C . ( 2014 ). The time course of spoken word recognition in Mandarin Chinese: A unimodal ERP study . Neuropsychologia , 63 , 165 – 174 . doi: 10.1016/j.neuropsychologia.2014.08.015 OpenUrl CrossRef PubMed ↵ Ito , A. , Gambi , C. , Pickering , M. J. , Fuellenbach , K. , & Husband , E. M . ( 2020 ). Prediction of phonological and gender information: An event-related potential study in Italian . Neuropsychologia , 136 , 107291 . doi: 10.1016/j.neuropsychologia.2019.107291 OpenUrl CrossRef ↵ Karunathilake , I. M. D. , Brodbeck , C. , Bhattasali , S. , Resnik , P. , & Simon , J. Z . ( 2025 ). Neural Dynamics of the Processing of Speech Features: Evidence for a Progression of Features from Acoustic to Sentential Processing . Journal of Neuroscience , 45 ( 11 ). doi: 10.1523/JNEUROSCI.1143-24.2025 OpenUrl Abstract / FREE Full Text ↵ Karunathilake , I. M. D. , Kulasingham , J. P. , & Simon , J. Z . ( 2023 ). Neural tracking measures of speech intelligibility: Manipulating intelligibility while keeping acoustics unchanged . Proceedings of the National Academy of Sciences , 120 ( 49 ), e2309166120 . doi: 10.1073/pnas.2309166120 OpenUrl CrossRef PubMed ↵ Keitel , A. , Gross , J. , & Kayser , C . ( 2018 ). Perceptually relevant speech tracking in auditory and motor cortex reflects distinct linguistic features . PLOS Biology , 16 ( 3 ), e2004473 . doi: 10.1371/journal.pbio.2004473 OpenUrl CrossRef PubMed Kuperberg , G. R. , & and Jaeger, T. F. ( 2016 ). What do we mean by prediction in language comprehension? Language, Cognition and Neuroscience , 31 ( 1 ), 32 – 59 . doi: 10.1080/23273798.2015.1102299 OpenUrl CrossRef PubMed ↵ Kutas , M. , & Federmeier , K. D . ( 2000 ). Electrophysiology reveals semantic memory use in language comprehension . Trends in Cognitive Sciences , 4 ( 12 ), 463 – 470 . doi: 10.1016/S1364-6613(00)01560-6 OpenUrl CrossRef PubMed Web of Science ↵ Kutas , M. , & Hillyard , S. A . ( 1980 ). Reading senseless sentences: Brain potentials reflect semantic incongruity . Science , 207 ( 4427 ), 203 – 205 . doi: 10.1126/science.7350657 OpenUrl Abstract / FREE Full Text ↵ Levy , R . ( 2008 ). Expectation-based syntactic comprehension . Cognition , 106 ( 3 ), 1126 – 1177 . doi: 10.1016/j.cognition.2007.05.006 OpenUrl CrossRef PubMed Web of Science ↵ Li , X. , Li , X. , & Qu , Q . ( 2022 ). Predicting phonology in language comprehension: Evidence from the visual world eye-tracking task in Mandarin Chinese . Journal of Experimental Psychology: Human Perception and Performance , 48 ( 5 ), 531 – 547 . doi: 10.1037/xhp0000999 OpenUrl CrossRef PubMed ↵ Li , X. , Rayner , K. , & Cave , K. R . ( 2009 ). On the segmentation of Chinese words during reading . Cognitive Psychology , 58 ( 4 ), 525 – 552 . doi: 10.1016/j.cogpsych.2009.02.003 OpenUrl CrossRef PubMed ↵ Linzen , T. , & Jaeger , T. F . ( 2016 ). Uncertainty and Expectation in Sentence Processing: Evidence From Subcategorization Distributions . Cognitive Science , 40 ( 6 ), 1382 – 1411 . doi: 10.1111/cogs.12274 OpenUrl CrossRef PubMed Liu , S. , Chen , X. , & Wang , S . ( 2025 ). The role of tonal information in speech prediction: Evidence based on Chinese tone sandhi . Language, Cognition and Neuroscience , 0 ( 0 ), 1 – 17 . doi: 10.1080/23273798.2025.2506637 OpenUrl CrossRef Liu , S. , Zhang , W. , & Wang , S . ( 2025 ). Neural Evidence for Tonal Prediction: Multivariate Decoding of Predicted Tone Categories Using Functional Magnetic Resonance Imaging Data . Journal of Cognitive Neuroscience , 1 – 16 . doi: 10.1162/jocn.a.84 OpenUrl CrossRef ↵ Mahowald , K. , Ivanova , A. A. , Blank , I. A. , Kanwisher , N. , Tenenbaum , J. B. , & Fedorenko , E . ( 2024 ). Dissociating language and thought in large language models . Trends in Cognitive Sciences , 28 ( 6 ), 517 – 540 . doi: 10.1016/j.tics.2024.01.011 OpenUrl CrossRef ↵ Marslen-Wilson , W . ( 1973 ). Linguistic Structure and Speech Shadowing at Very Short Latencies . Nature , 244 ( 5417 ), 522 – 523 . doi: 10.1038/244522a0 OpenUrl CrossRef PubMed Web of Science ↵ Mattys , S. L. , White , L. , & Melhorn , J. F . ( 2005 ). Integration of Multiple Speech Segmentation Cues: A Hierarchical Framework . Journal of Experimental Psychology: General , 134 ( 4 ), 477 – 500 . doi: 10.1037/0096-3445.134.4.477 OpenUrl CrossRef PubMed Web of Science ↵ McAuliffe , M. , Socolof , M. , Mihuc , S. , Wagner , M. , & Sonderegger , M . ( 2017 ). Montreal Forced Aligner: Trainable Text-Speech Alignment Using Kaldi . Interspeech 2017 , 498 – 502 . doi: 10.21437/Interspeech.2017-1386 OpenUrl CrossRef ↵ McDermott , J. H . ( 2009 ). The cocktail party problem . Current Biology , 19 ( 22 ), R1024 – R1027 . doi: 10.1016/j.cub.2009.09.005 OpenUrl CrossRef PubMed Web of Science ↵ Mesgarani , N. , Cheung , C. , Johnson , K. , & Chang , E. F . ( 2014 ). Phonetic feature encoding in human superior temporal gyrus. Science (New York , N.Y .) , 343 ( 6174 ), 1006 – 1010 . doi: 10.1126/science.1245994 OpenUrl Abstract / FREE Full Text ↵ Miller , J. , Patterson , T. , & Ulrich , R . ( 1998 ). Jackknife-based method for measuring LRP onset latency differences . Psychophysiology , 35 ( 1 ), 99 – 115 . doi: 10.1111/1469-8986.3510099 OpenUrl CrossRef PubMed Web of Science ↵ Näätänen , R. , Lehtokoski , A. , Lennes , M. , Cheour , M. , Huotilainen , M. , Iivonen , A. , Vainio , M. , Alku , P. , Ilmoniemi , R. J. , Luuk , A. , Allik , J. , Sinkkonen , J. , & Alho , K . ( 1997 ). Language-specific phoneme representations revealed by electric and magnetic brain responses . Nature , 385 ( 6615 ), 432 – 434 . doi: 10.1038/385432a0 OpenUrl CrossRef PubMed Web of Science ↵ Nelson , M. , Dehaene , S. , Pallier , C. , & Hale , J. ( 2017 ). Entropy Reduction correlates with temporal lobe activity . In T. Gibson , T. Linzen , A. Sayeed , M. van Schijndel , & W. Schuler (Eds.), Proceedings of the 7th Workshop on Cognitive Modeling and Computational Linguistics (CMCL 2017) (pp. 1 – 10 ). Association for Computational Linguistics . doi: 10.18653/v1/W17-0701 OpenUrl CrossRef ↵ Oostenveld , R. , Fries , P. , Maris , E. , & Schoffelen , J.-M . ( 2011 ). FieldTrip: Open Source Software for Advanced Analysis of MEG, EEG, and Invasive Electrophysiological Data . Computational Intelligence and Neuroscience , 2011 ( 1 ), 156869 . doi: 10.1155/2011/156869 OpenUrl CrossRef PubMed ↵ O’Seaghdha , P. G. , Chen , J.-Y. , & Chen , T.-M . ( 2010 ). Proximate units in word production: Phonological encoding begins with syllables in Mandarin Chinese but with segments in English . Cognition , 115 ( 2 ), 282 – 302 . doi: 10.1016/j.cognition.2010.01.001 OpenUrl CrossRef PubMed Web of Science ↵ Pan , D. J. , Yang , X. , Lui , K. F. H. , Lo , J. C. M. , McBride , C. , & Ho , C. S . ( 2021 ). Character and word reading in Chinese: Why and how they should be considered uniquely vis-à-vis literacy development . Contemporary Educational Psychology , 65 , 101961 . doi: 10.1016/j.cedpsych.2021.101961 OpenUrl CrossRef ↵ Pickering , M. J. , & Garrod , S . ( 2013 ). An integrated theory of language production and comprehension . Behavioral and Brain Sciences , 36 ( 4 ), 329 – 347 . doi: 10.1017/S0140525X12001495 OpenUrl CrossRef PubMed ↵ Qu , Q. , Damian , M. F. , & Kazanina , N . ( 2012 ). Sound-sized segments are significant for Mandarin speakers . Proceedings of the National Academy of Sciences , 109 ( 35 ), 14265 – 14270 . doi: 10.1073/pnas.1200632109 OpenUrl Abstract / FREE Full Text ↵ Qu , Q. , Feng , C. , Hou , F. , & Damian , M. F . ( 2020 ). Syllables and phonemes as planning units in Mandarin Chinese spoken word production: Evidence from ERPs . Neuropsychologia , 146 , 107559 . doi: 10.1016/j.neuropsychologia.2020.107559 OpenUrl CrossRef PubMed ↵ Radford , A. , Wu , J. , Child , R. , Luan , D. , Amodei , D. , & Sutskever , I . ( 2019 ). Language Models are Unsupervised Multitask Learners . ↵ Rouder , J. N. , Speckman , P. L. , Sun , D. , Morey , R. D. , & Iverson , G . ( 2009 ). Bayesian t tests for accepting and rejecting the null hypothesis . Psychonomic Bulletin & Review , 16 ( 2 ), 225 – 237 . doi: 10.3758/PBR.16.2.225 OpenUrl CrossRef PubMed ↵ Ryskin , R. , & Nieuwland , M. S . ( 2023 ). Prediction during language comprehension: What is next? Trends in Cognitive Sciences , 27 ( 11 ), 1032 – 1052 . doi: 10.1016/j.tics.2023.08.003 OpenUrl CrossRef PubMed ↵ Saberi , K. , & Perrott , D. R . ( 1999 ). Cognitive restoration of reversed speech . Nature , 398 ( 6730 ), 760 – 760 . doi: 10.1038/19652 OpenUrl CrossRef PubMed ↵ Sanders , L. D. , & Neville , H. J . ( 2000 ). Lexical, Syntactic, and Stress-Pattern Cues for Speech Segmentation. Journal of Speech , Language, and Hearing Research : JSLHR , 43 ( 6 ), 1301 – 1321 . doi: 10.1044/jslhr.4306.1301 OpenUrl CrossRef ↵ Sanders , L. D. , & Neville , H. J . ( 2003 ). An ERP study of continuous speech processing: I . Segmentation, semantics, and syntax in native speakers. Cognitive Brain Research , 15 ( 3 ), 228 – 240 . doi: 10.1016/S0926-6410(02)00195-7 OpenUrl CrossRef PubMed ↵ Sanders , L. D. , Newport , E. L. , & Neville , H. J . ( 2002 ). Segmenting nonsense: An event-related potential index of perceived onsets in continuous speech . Nature Neuroscience , 5 ( 7 ), 700 – 703 . doi: 10.1038/nn873 OpenUrl CrossRef PubMed Web of Science ↵ Sereno , J. A. , & Lee , H . ( 2015 ). The Contribution of Segmental and Tonal Information in Mandarin Spoken Word Processing . Language and Speech , 58 ( Pt 2 ), 131 – 151 . doi: 10.1177/0023830914522956 OpenUrl CrossRef ↵ Shain , C. , Meister , C. , Pimentel , T. , Cotterell , R. , & Levy , R . ( 2024 ). Large-scale evidence for logarithmic effects of word predictability on reading time . Proceedings of the National Academy of Sciences , 121 ( 10 ). doi: 10.1073/pnas.2307876121 OpenUrl CrossRef ↵ Shen , W. , Hyönä , J. , Wang , Y. , Hou , M. , & Zhao , J . ( 2021 ). The role of tonal information during spoken-word recognition in Chinese: Evidence from a printed-word eye-tracking study . Memory & Cognition , 49 ( 1 ), 181 – 192 . doi: 10.3758/s13421-020-01070-0 OpenUrl CrossRef PubMed ↵ Shu , H. , Peng , H. , & McBride-Chang , C . ( 2008 ). Phonological awareness in young Chinese children . Developmental Science , 11 ( 1 ), 171 – 181 . doi: 10.1111/j.1467-7687.2007.00654.x OpenUrl CrossRef PubMed Web of Science ↵ Siok , W. T. , & Fletcher , P . ( 2001 ). The role of phonological awareness and visual-orthographic skills in Chinese reading acquisition . Developmental Psychology , 37 ( 6 ), 886 – 899 . doi: 10.1037/0012-1649.37.6.886 OpenUrl CrossRef PubMed Web of Science ↵ Song , M. , Wang , J. , & Cai , Q . ( 2024 ). The unique contribution of uncertainty reduction during naturalistic language comprehension . Cortex , 181 , 12 – 25 . doi: 10.1016/j.cortex.2024.09.007 OpenUrl CrossRef PubMed ↵ Taylor , W. L . ( 1953 ). “Cloze procedure”: A new tool for measuring readability . Journalism Quarterly , 30 , 415 – 433 . OpenUrl Web of Science ↵ Tezcan , F. , Weissbart , H. , & Martin , A. E . ( 2023 ). A tradeoff between acoustic and linguistic feature encoding in spoken language comprehension . eLife , 12 , e82386 . doi: 10.7554/eLife.82386 OpenUrl CrossRef PubMed ↵ Vallat , R . ( 2018 ). Pingouin: Statistics in Python . Journal of Open Source Software , 3 ( 31 ), 1026 . doi: 10.21105/joss.01026 OpenUrl CrossRef ↵ Van Den Brink , D. , Brown , C. M. , & Hagoort , P. ( 2001 ). Electrophysiological Evidence for Early Contextual Influences during Spoken-Word Recognition: N200 Versus N400 Effects . Journal of Cognitive Neuroscience , 13 ( 7 ), 967 – 985 . doi: 10.1162/089892901753165872 OpenUrl CrossRef PubMed Web of Science ↵ Van Petten , C. , Coulson , S. , Rubin , S. , Plante , E. , & Parks , M. ( 1999 ). Time course of word identification and semantic integration in spoken language. Journal of Experimental Psychology: Learning , Memory, and Cognition , 25 ( 2 ), 394 – 417 . doi: 10.1037/0278-7393.25.2.394 OpenUrl CrossRef PubMed Web of Science ↵ Van Petten , C. , & Luka , B. J. ( 2012 ). Prediction during language comprehension: Benefits, costs, and ERP components . International Journal of Psychophysiology , 83 ( 2 ), 176 – 190 . doi: 10.1016/j.ijpsycho.2011.09.015 OpenUrl CrossRef PubMed Web of Science ↵ Venhuizen , N. J. , Crocker , M. W. , & Brouwer , H . ( 2019 ). Semantic Entropy in Language Comprehension . Entropy , 21 ( 12 ), Article 12 . doi: 10.3390/e21121159 OpenUrl CrossRef ↵ Wagenmakers , E.-J. , Marsman , M. , Jamil , T. , Ly , A. , Verhagen , J. , Love , J. , Selker , R. , Gronau , Q. F. , Šmíra , M. , Epskamp , S. , Matzke , D. , Rouder , J. N. , & Morey , R. D . ( 2018 ). Bayesian inference for psychology. Part I: Theoretical advantages and practical ramifications . Psychonomic Bulletin & Review , 25 ( 1 ), 35 – 57 . doi: 10.3758/s13423-017-1343-3 OpenUrl CrossRef PubMed ↵ Weissbart , H. , Kandylaki , K. D. , & Reichenbach , T . ( 2020 ). Cortical Tracking of Surprisal during Continuous Speech Comprehension . Journal of Cognitive Neuroscience , 32 ( 1 ), 155 – 166 . doi: 10.1162/jocn_a_01467 OpenUrl CrossRef PubMed ↵ Wiener , S. , & Turnbull , R . ( 2016 ). Constraints of Tones, Vowels and Consonants on Lexical Selection in Mandarin Chinese . Language and Speech , 59 ( 1 ), 59 – 82 . doi: 10.1177/0023830915578000 OpenUrl CrossRef ↵ Willems , R. M. , Frank , S. L. , Nijhof , A. D. , Hagoort , P. , & van den Bosch, A. ( 2016 ). Prediction During Natural Language Comprehension . Cerebral Cortex , 26 ( 6 ), 2506 – 2516 . doi: 10.1093/cercor/bhv075 OpenUrl CrossRef PubMed ↵ Wong , A. W.-K. , & Chen , H.-C . ( 2008 ). Processing segmental and prosodic information in Cantonese word production. Journal of Experimental Psychology: Learning , Memory, and Cognition , 34 ( 5 ), 1172 – 1190 . doi: 10.1037/a0013000 OpenUrl CrossRef PubMed Web of Science ↵ Wong , A. W.-K. , & Chen , H.-C . ( 2009 ). What are effective phonological units in Cantonese spoken word planning? Psychonomic Bulletin & Review , 16 ( 5 ), 888 – 892 . doi: 10.3758/PBR.16.5.888 OpenUrl CrossRef PubMed ↵ Xu , X. , Ji , C. , Li , T. , & Pickering , M. J . ( 2024 ). The prediction of segmental and tonal information in Mandarin Chinese: An eye-tracking investigation . Language, Cognition and Neuroscience , 0 ( 0 ), 1 – 15 . doi: 10.1080/23273798.2024.2395549 OpenUrl CrossRef ↵ Yang , Q. , & Chen , Y . ( 2022 ). Phonological competition in Mandarin spoken word recognition . Language, Cognition and Neuroscience , 37 ( 7 ), 820 – 843 . doi: 10.1080/23273798.2021.2024862 OpenUrl CrossRef ↵ Yu , S. , Gu , C. , Huang , K. , & Li , P . ( 2024 ). Predicting the next sentence (not word) in large language models: What model-brain alignment tells us about discourse comprehension . Science Advances , 10 ( 21 ), eadn7744 . doi: 10.1126/sciadv.adn7744 OpenUrl CrossRef PubMed ↵ Zhang , Y. , Frassinelli , D. , Tuomainen , J. , Skipper , J. I. , & Vigliocco , G . ( 2021 ). More than words: Word predictability, prosody, gesture and mouth movements in natural language comprehension . Proceedings of the Royal Society B: Biological Sciences , 288 ( 1955 ), 20210500 . doi: 10.1098/rspb.2021.0500 OpenUrl CrossRef ↵ Zhao , J. , Guo , J. , Zhou , F. , & Shu , H . ( 2011 ). Time course of Chinese monosyllabic spoken word recognition: Evidence from ERP analyses . Neuropsychologia , 49 ( 7 ), 1761 – 1770 . doi: 10.1016/j.neuropsychologia.2011.02.054 OpenUrl CrossRef PubMed ↵ Zhao , Z. , Chen , H. , Zhang , J. , Zhao , X. , Liu , T. , Lu , W. , Chen , X. , Deng , H. , Ju , Q. , & Du , X . ( 2019 ). UER: An Open-Source Toolkit for Pre-training Models (No . arXiv : 1909 . 05658 ). arXiv. doi: 10.48550/arXiv.1909.05658 OpenUrl CrossRef ↵ Zhao , Z. , Ding , J. , Wang , J. , Chen , Y. , & Li , X . ( 2024 ). The flexibility and representational nature of phonological prediction in listening comprehension: Evidence from the visual world paradigm . Language and Cognition , 16 ( 2 ), 481 – 504 . doi: 10.1017/langcog.2023.38 OpenUrl CrossRef ↵ Zhao , Z. , Li , Y. , Hou , C. , Zhao , J. , Tian , R. , Liu , W. , Chen , Y. , Sun , N. , Liu , H. , Mao , W. , Guo , H. , Guo , W. , Wu , T. , Zhu , T. , Shi , W. , Chen , C. , Huang , S. , Chen , S. , Liu , L. , … Yan , K . ( 2023 ). TencentPretrain: A Scalable and Flexible Toolkit for Pre-training Models of Different Modalities (No . arXiv : 2212 . 06385 ). arXiv. doi: 10.48550/arXiv.2212.06385 OpenUrl CrossRef View the discussion thread. Back to top Previous Next Posted November 25, 2025. Download PDF Email Thank you for your interest in spreading the word about bioRxiv. NOTE: Your email address is requested solely to identify you as the sender of this article. Your Email * Your Name * Send To * Enter multiple addresses on separate lines or separate them with commas. You are going to email the following Predictive Neural Signals during Natural Mandarin Speech Comprehension Message Subject (Your Name) has forwarded a page to you from bioRxiv Message Body (Your Name) thought you would like to see this page from the bioRxiv website. Your Personal Message CAPTCHA This question is for testing whether or not you are a human visitor and to prevent automated spam submissions. Share Predictive Neural Signals during Natural Mandarin Speech Comprehension Qifei Wang , Eva Berlot , Judit Fazekas , Jakub Szewczyk , Floris P de Lange bioRxiv 2025.11.23.690006; doi: https://doi.org/10.1101/2025.11.23.690006 Share This Article: Copy Citation Tools Predictive Neural Signals during Natural Mandarin Speech Comprehension Qifei Wang , Eva Berlot , Judit Fazekas , Jakub Szewczyk , Floris P de Lange bioRxiv 2025.11.23.690006; doi: https://doi.org/10.1101/2025.11.23.690006 Citation Manager Formats BibTeX Bookends EasyBib EndNote (tagged) EndNote 8 (xml) Medlars Mendeley Papers RefWorks Tagged Ref Manager RIS Zotero Tweet Widget Facebook Like Google Plus One Subject Area Neuroscience Subject Areas All Articles Animal Behavior and Cognition (7635) Biochemistry (17691) Bioengineering (13892) Bioinformatics (41936) Biophysics (21452) Cancer Biology (18588) Cell Biology (25504) Clinical Trials (138) Developmental Biology (13378) Ecology (19899) Epidemiology (2067) Evolutionary Biology (24320) Genetics (15609) Genomics (22506) Immunology (17736) Microbiology (40394) Molecular Biology (17181) Neuroscience (88605) Paleontology (666) Pathology (2832) Pharmacology and Toxicology (4824) Physiology (7641) Plant Biology (15153) Scientific Communication and Education (2045) Synthetic Biology (4294) Systems Biology (9825) Zoology (2271)