Evoked category representations

Evoked category representations | Authorea try { document.documentElement.classList.add('js'); } catch (e) { } var _gaq = _gaq || []; _gaq.push(['_setAccount', 'G-8VDV14Y67G']); _gaq.push(['_trackPageview']); (function() { var ga = document.createElement('script'); ga.type = 'text/javascript'; ga.async = true; ga.src = ('https:' == document.location.protocol ? 'https://ssl' : 'http://www') + '.google-analytics.com/ga.js'; var s = document.getElementsByTagName('script')[0]; s.parentNode.insertBefore(ga, s); })(); Skip to main content Preprints Collections Wiley Open Research IET Open Research Ecological Society of Japan All Collections About About Authorea FAQs Contact Us Quick Search anywhere Search for preprint articles, keywords, etc. Search Search ADVANCED SEARCH SCROLL European Journal of Neuroscience This is a preprint and has not been peer reviewed. Data may be preliminary. 22 April 2025 V1 Latest version Share on Evoked category representations Authors : Chao Han 0000-0003-2867-3998 [email protected] , Arild Hestvik 0000-0003-4561-7584 , and W Idsardi Authors Info & Affiliations https://doi.org/10.22541/au.174529283.31063398/v1 337 views 149 downloads Contents Abstract Introduction Underspecification MMN asymmetries Cross-categorical phonetic distance effects The current study: Within-category contrasts Experiment 1 EEG recording and preprocessing Results Temporospatial PCA decomposition Factor score analysis Discussion Experiment 2 EEG recording and preprocessing Results Factor score analysis Discussion Experiment 3 Factor score analysis Discussion General Discussion Data Availability Statement Author contribution Acknowledgements Funding Information References Information & Authors Metrics & Citations View Options References Figures Tables Media Share Abstract A critical part of the Mismatch Negativity (MMN) mechanism is the construction of a memory trace encoding regularities extracted from the stimuli in an oddball paradigm. In an influential study, Phillips et al. (2000) argued that varying phonetic standards within the limits of a phoneme category prompts the auditory cortex to access representations of the “discrete phonological categories” (p. 1050), resulting in a mismatch response by comparing the deviant stimulus to an abstract, evoked category representation. The present study tested the strongest interpretation of this claim–namely, that the phoneme itself is retrieved from long-term memory and serves as the memory trace for the stimulus sequence. This interpretation has been the implicit, if not explicit, basis for a body of research employing the varying-standards paradigm to probe phonological underspecification. However, while previous research has focused on contrasts between distinct phonemes, we examined a previously untested prediction of this interpretation: that varying the standards should eliminate the MMN when the contrast is within a single phoneme category. Contrary to this prediction, we observed a mismatch negativity to a within-category deviant even when standards were varied. In two additional experiments we further examined whether the within-category MMN is from long-term memory representations of phonetic realizations of the phoneme, or from listeners constructing ad hoc statistical representations of the stimulus distribution. The weight of the evidence suggests that the within-category MMN observed with varying standards reflects sensitivity to the statistical structure of the stimuli, rather than activation of abstract phonological categories. Introduction Speech perception is the process of mapping continuous and variable speech signals to discrete linguistic phoneme categories (Pisoni & Luce, 1987). Speech signals describe articulatory and acoustic events in real time and space, while a phoneme category is an internal representation that serve as access codes to words (Kazanina et al., 2018). For example, the phoneme /t/ in American English can be phonetically realized with a voice onset time (VOT) ranging from 15 ms (a phonetic realization close to a [d] sound) to 160 ms (resembling an aspirated [t h ]) at the onset position of a stressed syllable (Chodroff & Wilson, 2018).11We use slashes (//) to denote phonemes and squared brackets ([]) to denote actual phonetic realizations, following the Handbook of the International Phonetic Association (1999). It can also be realized as a flap between two vowels (as in butter ) or as a glottal stop word-finally (as in hat trick ). Despite the acoustic variability in how phonemes are realized as speech signals, listeners have no difficulty mapping the variable realizations back to the discrete invariant phoneme /t/. Mapping speech signals to phonemes must involve knowledge of how a phonological category is phonetically realized (Kingston & Diehl, 1994), for example, by inspecting continuous variables such as VOT, where a longer VOT increases the likelihood that a given sound is mapped to /t/ rather than /d/. Both the bottom-up analysis of the acoustic signals as well as the stored knowledge of speech sound categories and their relationship to acoustic and articulatory events (top-down information) contribute to solving the mapping between external signals and internal mental representations. The Mismatch Negativity (MMN) event-related brain response (Näätänen et al., 1978) has long been used to investigate which properties of acoustic signals, as well as knowledge of native language speech sound categories, are utilized to detect differences between speech sounds. The MMN is an event-related potential (ERP) component that reflects the detection of an unexpected stimulus. It typically peaks within 150–300 ms after the onset of a “deviant” stimulus which breaks a regularity established by a sequence of “standard” tokens (Näätänen, 1990). In its scalp topography, the MMN exhibits a frontocentral negativity, with a positive polarity at mastoids when using an average reference. The presence of an MMN indicates that the auditory cortex differentiates between two stimulus types, whether the difference is related to pre-existing mental representations (such as sound categories) or to non-linguistic sensory differences (such as intensity differences for the same sound category). The effect of linguistic knowledge on sound signal differentiation has been thoroughly demonstrated in studies involving change detection among consonants (Sharma et al., 1993; Dehaene-Lambertz, 1997; Allen et al., 2000; Joanisse et al., 2007), vowels (Näätänen et al., 1997; Winkler et al., 1999; Shestakova et al., 2002; Sittiprapaporn et al., 2005; Miglietta et al., 2013), as well as complex spectro-temporal patterns (Näätänen et al., 1993; Schröger et al., 1994). Earlier literature often did not differentiate between phonetic versus phonemic distinctions. When comparing two stimuli from two different phoneme categories (i.e., an across-category contrast), such as a deviant [d] versus a standard [t], the categorical contrast is necessarily accompanied by a phonetic/acoustic contrast. An observed mismatch can therefore be attributed to either an acoustic difference between two speech sounds, or a phoneme category difference, or both. Indeed, Näätänen et al. (1997) observed a larger MMN to a cross-category contrast compared to a within-category contrast (where both standards and deviants belong to the same phoneme category), and attributed the within-category MMN to “the (mere) acoustic deviance” (p. 433). This raises the question: Is it possible to observe a mismatch response due to the phoneme-category difference alone, without reference to the accompanying physical contrast? Phillips et al. (2000) answered this question in the affirmative, by employing the “varying-standards paradigm”: Instead of using a single token stimulus for all standards (hereafter as “single-standard paradigm”), the varying-standards paradigm varied standards along some acoustic parameter, e.g., VOT, such that no single standard could serve as the memory trace. This manipulation still resulted in an MMN, suggesting that listeners were able to construct an abstract memory representation based on the linguistic knowledge that all standards were drawn from a single phoneme category. Importantly, when the possibility of grouping stimuli based on abstract category knowledge was disturbed, the MMN vanished. In a second experiment in Phillips et al. (2000), the VOT values of all the stimuli were shifted up by 20 ms, such that the frequently occurring “standards” now straddled the /d/ and /t/ categories. The total set of stimuli still preserved the many-to-few ratio based on VOT–there were more frequent stimuli with relatively shorter VOTs and fewer stimuli with relatively longer VOTs–yet no MMN was observed. This suggests that listeners did not sort the stimuli into phonetically based ad hoc categories of “many relatively short VOT stimuli” versus “few relatively long VOT stimuli.” Phillips et al. (2000) concluded that the mismatch effect in their first experiment arose because “the left-hemisphere auditory cortex has access to representations of discrete phonological categories” (p. 1051), which allowed listeners to categorize the more frequent relatively shorter VOT stimuli as /d/, and the oddball long VOT stimuli as /t/. In the “shifted” condition, no single phoneme category could be used as the memory trace, which caused the absence of a mismatch effect. The influential argument arising from this finding was that by varying the standards (within the phoneme category limits), the MMN mechanism accesses the abstract phoneme. As a result, the MMN arises from a comparison of the abstract phoneme with the phonetic deviant to generate a mismatch effect. In the next section, we spell out a linking theory by formalizing the MMN computation as unification (Shieber, 1986) process. Linking MMN and phonological category The ingredients for an MMN are (a) a memory trace representation of or derived from standards, (b) a representation of the deviant stimulus, and (c) a comparison process.22The current study adopts a memory-based interpretation of the MMN mechanism. However, it should be acknowledged that the memory-based model is but one theory of MMN, cf., the adaptation-based model (e.g., May & Tiitinen, 2010; May, 2021) and the predictive coding model (e.g., Garrido et al., 2009). Consider first the memory trace in an MMN experiment comparing two different speech sound tokens, for example repeated single-token standards [d] with a 20-ms VOT versus a deviant [t] with a 60-ms VOT. In this case, the memory trace is simply an echoic memory representation of the repeated standards, with values for continuous variables such as intensity, frequency, amplitude, duration, VOT, frication noise, etc. An auditory system without an associated phonological system (say, that of a chinchilla) would therefore be expected to generate a mismatch response purely based on the comparison of continuous VOT values. If we add phonological features to the sound representations (as humans would have), the standard and deviant would not only differ in terms of VOT, but also in terms of phonological features like [+/-voice].33We use traditional binary features ([+/-]) to specify the presence or absence of a feature value, which is sufficient for the scope of the current study. It should be noted that the phonological underspecification discussed below employs a monovalent feature approach (Lahiri & Reetz, 2010). Two or more identical [d] stimuli with the same VOT value would incrementally lead to a memory trace of the sound with the given VOT value. When a new stimulus that differs only in VOT is encountered, the difference in VOT value generates a mismatch signal. What happens with the memory trace when standards are varied? As Phillips et al. state, no echoic memory representation of a single stimulus can now alone be the memory trace. If the auditory cortex now can “access” the phoneme, the simplest model of this effect is to say that the “accessed phoneme” is retrieved from long-term memory and used as the de facto memory trace representation of the varying standards. More complex models can be envisaged, such as a memory trace containing both the phoneme as well as acoustic information, but the goal of the current study is to ascertain the simplest model of “the phoneme as memory trace”. Since the phoneme, per theory, is a discrete, symbolic representation, it no longer contains continuous variables such as VOT. How can then a comparison be carried out between the memory trace and an acoustic oddball stimulus that is drawn from a different phoneme category? The degree to which a deviant is indeed a deviant can now only be calculated by comparing the standard and deviant on the same shared set of variables that define both objects. According to linguistic theory, this is the set of discretely valued “phonological features”. Specifically, no mismatch can be defined in terms of continuous variables like VOT, because this information is now missing from the phoneme memory trace representation. The comparison between memory trace and deviant representation must be in terms of the variables that phonemes share with phonetic units–the phonological features (Chomsky & Halle, 1968). Following linguistic theory, we assume that the phonetic representation of the oddball is a superset of the features that define the phoneme. For concreteness, we can formalize the MMN computation process as unification (Shieber, 1986), and the MMN itself as the consequence of unification failure. A memory trace (now, a bundle of distinctive features defining a phoneme) is compared to the next stimulus in a sequence via an incremental process of feature unification: Every new stimulus is sought to be unified with the preceding set. A mismatching stimulus can then be defined as a feature set containing values that are contradictory to those in the preceding set. In the case of a single VOT standard as discussed above, this will lead to unification failure due to two distinct VOT values as well as two contradictory voicing feature values (Example 1). [t] [d] \(\begin{bmatrix}-voice\\ -\text{continuant}\\ \text{POA}:\text{coronal}\\ VOT:60\\ \ldots\\ \end{bmatrix}\) ⨆ \(\begin{bmatrix}+voice\\ -continuant\\ POA:coronal\\ VOT:20\\ \ldots\\ \end{bmatrix}\) = Unification failure However, when standards are varied and the phoneme is used as the memory trace, the comparison of, say, /t/ when compared to an oddball [d], leads to unification failure solely due to two contradictory voicing feature values, as opposed to distinct VOT values (Example 2). /t/ [d] \(\begin{bmatrix}-voice\\ -\text{continuant}\\ \text{POA}:\text{coronal}\\ \ldots\\ \end{bmatrix}\) ⨆ \(\begin{bmatrix}+voice\\ -continuant\\ POA:coronal\\ VOT:20\\ \ldots\\ \end{bmatrix}\) = Unification failure It is now straightforward to model the generation of a mismatch signal based on any contradictory feature value from any domain. In the case of a single-token standard, the mismatch signal is generated by contradictory acoustic details, e.g., VOT. If we add the cognitive system of a human, we also have contradictory phonological feature values, i.e., [+voice] and [-voice]. When standards are varied, the difference is that mismatch is then generated solely on the basis of contradictory values of discrete, phonological variables, e.g., [+voice] and [-voice]. Such a comparison process has formed the basis for a long line of research on MMN evidence for phonological underspecification (Lahiri & Reetz, 2002; Eulitz & Lahiri, 2004; Hestvik & Durvasula, 2016; Scharinger et al., 2016; Monahan, 2018; Hestvik et al., 2020, and many others). We review the experimental logic of this research next, as it serves to highlight the innovativeness of the varying-standards linking model. Underspecification MMN asymmetries Underspecification means that in minimal pairs of phonemes, one member of the pair has no specification for a given distinctive feature. Underspecification only applies to phonemes, while phonetic representations are always fully specified, as required by the principle of Full Interpretation (Chomsky, 1986). If, for example, /d/ is underspecified for the phonological feature [+voice] and functions as a standard, feature unification will succeed, and no mismatch effect is expected (Example 3). /d/ [t] \(\begin{bmatrix}\varnothing\\ -\text{continuant}\\ \text{POA}:\text{coronal}\\ \ldots\\ \end{bmatrix}\) ⨆ \(\begin{bmatrix}-voice\\ -continuant\\ POA:coronal\\ VOT:60\\ \ldots\\ \end{bmatrix}\) = \(\begin{bmatrix}-voice\\ -continuant\\ POA:coronal\\ VOT:60\\ \ldots\\ \end{bmatrix}\) On the other hand, if the standard is /t/, the phoneme memory trace does have a [-voice] feature specification. A phonetic oddball [d], as phonetic representations are always fully specified, will then contain the feature [+voice] and a feature clash ensues, as in Example 2 above. This predicts an MMN for oddball [d] when compared to varying /t/ standards, but no MMN for [t] compared to varying /d/ standards (Hestvik & Durvasula, 2016; Schluter et al., 2017). This logic of MMN asymmetry predictions for underspecified phonemes has been applied to a range of contrasts and languages. Notably, this logic only goes through if the memory trace indeed is a phoneme representation, and the comparison process determining a match or mismatch relies on phonological computation (i.e., the computation here is determining equality between discrete feature values). The observed underspecification-related MMN asymmetries in experiments using varying standards can therefore be seen as a success for the claim that varying the standards provide access to the abstract phoneme representations in long-term memory and provided a method for assessing the nature of these abstract representations. In addition to MMN asymmetries when underspecified phonemes are compared to specified phonetic oddballs, the varying-standards linking theory also makes predictions about a lack of phonetic distance effects when different cross-category deviants are compared to a phoneme, which we discuss in the next section. Cross-categorical phonetic distance effects MMN in a single-standard, sensory based experiment typically shows a greater amplitude MMN the greater the relevant physical or psychophysical distance is between standards and deviants (e.g., Näätänen et al., 2007). However, if varying the standards results in a phoneme memory trace, the prediction is that the mismatch generator should be indifferent to the gradient phonetic characteristic of either the standards or the deviants, and amplitude modulations based on varying physical differences between standards and deviants should disappear. The phonetic characteristics of the stimuli that triggered the retrieval of the phoneme do not come along with the abstract phoneme. This predicts that the acoustic distance between a deviant and the varying standards, in a cross-categorical experiment, should not have an effect on the MMN. This prediction was tested in Rhodes et al. (2019). The study had participants listen to synthesized CV syllables consisting of an alveolar stop followed by the vowel /æ/ as stimuli. The VOT of those syllables were drawn from a /d–t/ continuum. The standards were always voiceless alveolar [t] stops, but the standards were presented in two conditions: A “low” condition with VOT values of 60, 65, and 70 ms, and a “high” condition with higher VOT values of 75, 80, and 85 ms. The deviant stimulus in both conditions was a 15-ms VOT [d]. If the phonological memory trace erases information about the acoustic details of the proximal standard stimuli presented in the experiment, the deviant should elicit the same MMN response when embedded in standards with higher VOTs as it does in standards with lower VOTs. On the other hand, if the memory trace retained the VOT information of the varying standards, then an MMN difference should be observed. No difference in MMN amplitude was observed. This is consistent with the linking theory that the memory trace within the varying-standards paradigm exclusively contains discrete phonological features. The memory trace is a symbolic, discrete, abstract phoneme representation, which means that it only contains a distinctive feature matrix, which per theory (Chomsky & Halle, 1968) does not contain gradient phonetic information. Consequently, there can be no comparison effects “across representational domains”, for example between a phoneme and an oddball that is defined in terms of its analog, phonetic, or gradient information, because phonemes (in theory) are not defined for this type of information. Rhodes et al. (2022) also report similar results. In summary, the underspecification literature and the studies of Rhodes et al. (2019, 2022), comparing cross-categorical contrasts with varying standards, provide evidence that varying the standards generate MMN based on categorical differences alone. The current study: Within-category contrasts However, the prediction of the above linking theory, which to our knowledge has not been directly tested with a dedicated experiment, is that a within-category deviant should not result in an MMN if standards are varied. Consider the scenario where a phoneme memory trace for /d/ is set up by varying the VOT values of the standards. When an incrementally constructed phoneme memory trace is compared to any instance of an allophone of /d/, say, a [d] with a 20-ms VOT, unification will succeed because the phoneme’s feature set is a proper subset of the phonetic deviant:11We represent the phoneme /d/ here underspecified for voicing, to be consistent with the preceding discussion. However, the prediction that varying standards should eliminate within-category MMN is independent of underspecification. /d/ [d] \(\begin{bmatrix}\varnothing\\ -\text{continuant}\\ \text{POA}:\text{coronal}\\ \ldots\\ \end{bmatrix}\) ⨆ \(\begin{bmatrix}+voice\\ -continuant\\ POA:coronal\\ VOT:20\\ \ldots\\ \end{bmatrix}\) = \(\begin{bmatrix}+voice\\ -continuant\\ POA:coronal\\ VOT:20\\ \ldots\\ \end{bmatrix}\) On the other hand, if the /d/ standards are not varied but a single [d] with a 0 ms VOT serves as the standard, then by our linking theory, an echoic, phonetic memory trace is built up. In this case, the continuous VOT variable is encoded in the memory trace, and when compared to a 20-ms VOT [d], unification would fail, and a mismatch signal should be generated. The two predictions regarding varying versus single standards were indirectly supported by findings from Kazanina et al. (2018) and Hestvik and Durvasula (2016). Kazanina et al. presented standards and deviants with varying VOTs which were categorized as distinct phonemes (/d/ and /t/) in Russian, but as allophonic variants of /t/ in Korean. They found that Russian speakers exhibited a mismatch effect while no MMN was elicited in Korean speakers. Hestvik and Durvasula observed an MMN asymmetry in a varying-standards experiment comparing /d/ and /t/ in American English, which was explained in terms of phonological underspecification. They also reported an experiment where standards were not varied but the same contrast was tested. As predicted, the asymmetry related to abstract feature comparisons now disappeared and the same MMN was observed for both contrasts (standard /d/ versus deviant [t], and standard /t/ versus deviant [d]). This outcome is consistent with the idea that single-token standards lead to the encoding of phonetic/acoustic details in the memory trace. While the above two study provide evidence that within-category MMN should disappear with varied standards but emerge with single-token standards, it remains unclear whether the absent of an MMN effect in the varying-standards condition was due to insufficient acoustic distance between standards and deviants. Such a limitation could prevent auditory stream segregation–a prerequisite for deviance detection (see Sussman et al., 1999; Sussman et al., 2014). Therefore, it is necessary to compare within-category MMN responses elicited by varying versus single-token standards using the same acoustic difference to the deviant, within the same participants (and therefore a fortiori within-language). We here aim to fill this gap. Experiment 1 investigates the central prediction by comparing a single-standard condition (in which a within-category MMN is expected) to a varying-standards condition (in which the MMN is predicted to disappear). To preview the results, MMN responses were observed in both conditions. Experiment 2 and 3 were then conducted to investigate the source of this within-category MMN–specifically, whether it arises from a statistical tracking and representation of the VOT values of the standards, or it arise from activation of a phonetic prototype category. Experiment 1 Experiment 1 tested the prediction that when varying the standards, an MMN should not be elicited by a within-category deviant, even if this deviant has physical characteristics that make it different from the set of varying standards. We used the voiceless alveolar stop /t/ as the test phoneme for this prediction, by presenting the stimulus with three different VOT values (mean = 48 ms) as standards, and that with a 119-ms VOT as deviants. Even though the within-category deviant is physically distant from the standards, no MMN is expected if varying the standards leads to the construction of a symbolic phoneme /t/ as memory trace. Since the central prediction here is a null effect, which is difficult to interpret in isolation, we added a single-standard condition as a control, where an MMN response is always expected. The single-standard condition used a single standard stimulus with a 48-ms VOT combined with the same deviant stimulus with a 119-ms VOT. The statistical outcome supporting a memory trace of a symbolic phoneme is that the single-standard condition results in a significant MMN whereas the varying-standards condition does not. Another way of framing this prediction is to say that a phonetic/acoustic MMN (e.g., Miglietta et al., 2013) should be manifested with the single-standard condition, but disappear with the varying-standards condition. Methods Participants and stimuli Sixty-three native speakers of English, aged 18-30 (55 females, mean age = 21), were recruited from the University of Delaware. All participants reported no history of language impairment. Participants received either $20 or extra course credits for completing the experiment. The experimental procedure was approved by the University of Delaware Internal Review Board and was compliant with the principles for ethical research established by the Declaration of Helsinki. The stimuli were a set of CV syllables composed of an alveolar stop and the vowel /æ/. The alveolar stop was drawn from a /d–t/ continuum resynthesized in Praat (Winn, 2020; Boersma & Weenink, 2022). The voice onset time (VOT) of the continuum ranged from 0 to 145 ms, increasing by 1 ms at each step. All stimuli had a duration of 620 ms and an F0 onset frequency of 279 Hz. To mitigate abrupt changes and potential clicks at the beginning and end of each audio signal, a cosine ramp was applied during the first 5 and the last 80 ms. In addition, two non-speech stimuli were created by replacing a portion of the 48-ms VOT stimulus with a pure tone of 440 Hz. The replaced portion was either the first 150 ms following the vowel onset or the last 150 ms of the stimulus. Design and procedure The experimental design consisted of three blocks: a single-standard oddball block, a varying-standards oddball block, and a roving-standards control block (Figure 1). In both oddball blocks, only syllables with /t/ onset were presented. In the single-standard oddball block, /tæ/ tokens with a shorter VOT (48 ms) served as standards, while those with a longer VOT (119 ms) served as deviants. In the varying-standards oddball block, the VOT of standards varied between 42, 48, and 55 ms, while the VOT of deviants remained the same 119 ms. The roving-standards control block alternated between sequences of /dæ/ tokens with a 19-ms VOT and sequences of /tæ/ tokens with a 119-ms VOT. Starting from the second sequence, the initial token within each sequence was treated as a deviant, and the following tokens as standards. Each block contained 100 deviant tokens, with a standard-to-deviant ratio of 6:1, and an interstimulus interval ranging from 650 to 920 ms in increments of 10 ms. The number of successive standards preceding a deviant ranged from 3 to 9. Note that in the current design, the within-category MMN can be derived by comparing the deviant /t/ (VOT: 119 ms) to either the 48-ms VOT standard /t/ within the same oddball block (i.e., a within-block MMN) or to the 119-ms VOT standard /t/ from the roving-standards control block (i.e., an identity MMN). We chose the latter approach because it allowed us to compute an MMN using the identical stimulus, thereby eliminating the influence of different physical properties (Jacobsen et al., 2003). For clarity, we use the term ”control” to refer to the 119-ms VOT standard /t/ from the roving-standards control block. The experiment was programmed and presented using the E-Prime 2.0 software. Participants were assigned in a counterbalanced manner to either a single-standard group ( n = 31) or a varying-standards group ( n = 32). Each session began with the roving-standards control block, followed by the single-standard oddball block for participants in the single-standard group, or the varying-standards oddball block for those in the varying-standards group. Each session contained 67 randomly interspersed target tokens which were randomly sampled from the two non-speech stimuli. Participants were instructed to press a button upon detecting a non-speech sound. An image of a jigsaw puzzle piece then appeared on the screen, colored if the button was pressed within 1.5 seconds or in greyscale otherwise (Figure 1). The inclusion of the non-speech targets and the jigsaw puzzle game was intended to prevent participant boredom. The mean detection rate of the non-speech targets exceeded 99% for both groups, with 60 out of 63 participants detecting 100% of the non-speech targets. ­­­ Figure : The schematic representation of Experiment 1 design. Numbers are VOT values. Standards are highlighted in blue and deviants in red. Musical notes indicate non-speech targets. Double-sided arrows and rounded rectangles indicate the contrasts to derive an identity MMN for the 119-ms VOT. To confirm that all tokens presented in the oddball blocks were perceived as /tæ/—thus constituting a within-category contrast—we conducted a two-alternative forced-choice phoneme identification task after the electroencephalogram (EEG) recording to determine each participant’s /d–t/ perceptual boundary. The task required participants to judge whether a presented CV syllable began with /d/ or /t/, with VOT values sampled from 0 to 70 ms in 5-ms increments, each presented six times. EEG recording and preprocessing The continuous EEG data were recorded using EGI’s 64-channel HydroCel Geodesic Sensor Nets and the Net Station 4.5 software running on an EGI 300 system, with a sampling rate of 250 Hz. The impedance of each electrode was lowered to below 50 kΩ before the recording. The Channel E65 (corresponding to Cz in the 10-10 system) served as the reference channel. The entire experiment session took about 1.5 hours, including EEG net placement, instructions, breaks, net removal, and the subsequent phoneme identification task. The collected data were filtered using an offline 0.1–40 Hz bandpass filter. The filtered data were then segmented into epochs of 1000 ms, time-locked to the stimulus onset. Each epoch contained a pre-stimulus window of 200 ms, which was used for baseline correction. The epochs underwent an Multi-Algorithm Artifact Correction (MAAC) procedure to remove bad channels and movement artifacts, using the ERP PCA toolkit (Dien, 2010, 2024). For each epoch, a channel was marked bad if its highest absolute correlation with its neighboring channels fell below r = .4 across all time points. Bad channels were replaced via a spline interpolation from surrounding good channels. If a channel was marked bad in over 20% of epochs, it was considered a globally bad channel. Any epoch with more than 6 bad channels was excluded. The bad channel corrected data were then submitted to an automated eyeblink subtraction, which calls the ‘runica‘ function from EEGLAB (Delorme & Makeig, 2004). If an ICA component was correlated at r = .9 or greater with the automatically generated eyeblink template, it was marked as an eyeblink component and subtracted from the data. The resulting data were average re-referenced. For reproducibility, the data, analysis, and results are available at the Open Science Framework repository (https://osf.io/m5kr9/?view_only=fcd3fce78c08449ca8d15badb651c7c0). Results Phoneme identification task For each participant, a logistic regression was fitted to model their responses to the VOT continuum. The /d–t/ perceptual boundary was defined as the VOT value corresponding to a 50% probability of /t/ responses. The mean VOT boundary averaged across all 63 participants was 34 ms (SD = 4.5, range = 25–48) (Figure 2A). To ensure that all participants included in the ERP analysis perceived all tokens presented in the oddball blocks (VOT: 42, 48, 55, and 119 ms) as phoneme /t/, participants with a VOT boundary above 40 ms were excluded (Figure 2B). As a result, four participants from the single-standard group and one from the varying-standards group were excluded. The remaining 58 participants were included in the EEG data analysis. Figure : A: Proportion of /t/ responses at each VOT value, averaged across all 63 participants. Error bars indicate standard errors. The dashed vertical line indicates the mean perceptual boundary (34 ms). B: Distribution of individual perceptual boundaries. The dashed vertical line indicates the exclusion cutoff of 40 ms. Temporospatial PCA decomposition To identify the temporal dynamic and spatial distribution of the MMN effect, we conducted a sequential temporospatial Principal Component Analysis (PCA) using the ERP PCA toolkit (Dien, 2010; Luck & Gaspelin, 2017). The sequential temporospatial PCA procedure consists of an initial temporal PCA, which decomposes the observed ERPs into latent temporal factors, followed by a spatial Independent Component Analysis (ICA), which decomposes each temporal factor into latent spatial factors. For the input to the temporospatial PCA, we averaged the brain responses to obtain the ERP of the control /t/ (VOT: 119 ms) from the roving-standards control block and the ERP of the deviant /t/ (VOT: 119 ms) from the oddball block for each participant. The temporal PCA was performed on the covariance matrix computed from the input data, in which time points were arranged as columns and participants, electrodes, and conditions (i.e., control and deviant) as rows. Singular value decomposition was used for factor extraction, with Kaiser weighting and Promax rotation (κ = 3). To determine the number of factors to extract, we followed the parallel analysis approach by comparing the scree plot from a PCA retaining as many factors as there were time points to the scree plot of a PCA run on a dataset with permuted time points (Horn, 1965; Dien, 1998). The number of factors was determined up to the point at which the two scree plots intersected, yielding 11 factors. The temporal PCA was then rerun to extract the 11 temporal factors, which accounted for 95% of the total variance. Figure 3A illustrates the unstandardized factor loadings of all extracted temporal factors, along with their corresponding explained variance and peak latencies. Each temporal factor was subsequently submitted to a spatial ICA, using the same parameters as the temporal PCA, except with Infomax rotation. The spatial ICA extracted four temporospatial factors (TFSFs) from each temporal factor (TF), with the number determined based on the average scree plot across all 11 temporal factors, following the same parallel analysis approach described above. The retained temporospatial factors altogether account for about 72% of the total variance. Reconstructed ERPs were computed at each electrode for each temporal factor and temporospatial factor by first averaging the factor scores across participants for each stimulus condition (i.e., control and deviant). The averaged scores at each electrode were then multiplied by the temporal factor loading at each time point to obtain the reconstructed ERP waveform for temporal factors, and were additionally multiplied by the corresponding spatial factor loading at that electrode to obtain the reconstructed ERP waveform for temporospatial factors. We then computed reconstructed difference ERPs for each participant by subtracting the reconstructed deviant waveforms from the reconstructed control waveforms. We selected TF5SF2 (peak latency: 236 ms, electrode with maximal loading: E65)—which peaked within the typical MMN time window and exhibited a central negativity in the topographies of the grand-average reconstructed difference ERPs—as indicative of the MMN effect. Figure 3B shows topographies of the grand-average reconstructed difference ERPs for TF5 and its four temporospatial factors. Figure : Temporal factor loadings and temporospatial factor topographies from Experiment 1. A: Unstandardized factor loadings of the 11 temporal factors. Legends show peak latencies and explained variance. The thick line highlights TF5. B: Topographies of grand-average reconstructed difference ERPs (reconstructed deviant /t/ minus reconstructed control /t/) at the peak latency, for TF5 and its four temporospatial factors. Figure 4 illustrates the grand-average observed ERPs overlayed with the TF5SF2-reconstructed ERPs at the maximum electrode E65 for each group. The reconstructed ERPs depicted the contribution of the temporospatial factor to the observed ERPs. Figure : Grand-average observed ERPs overlaid with reconstructed ERPs at the maximum electrode for the single-standard group (left panel) and the varying-standards group (right panel). The lower panel shows difference waves. Factor score analysis The MMN amplitude was measured using the factor scores derived from TF5SF2, which peaked at 236 ms and exhibited a central topography. Those factor scores quantified the degree to which each participant and each stimulus condition contributed to TF5SF2, as determined by its temporal and spatial factor loadings. Because the factor loadings already quantified the contribution of each time point and electrode to the factor, it was unnecessary to preselect a specific time window or spatial region of interest. This approach minimizes subjectivity (i.e., preselecting an analysis time window and electrode region) and mitigates the multiple comparisons problem inherent in mass univariate analyses across time and space. Accordingly, a significant MMN effect would be reflected in a significant difference between the factor scores for the control versus the deviant, with higher scores for the former and lower scores for the latter. The factor scores were submitted to a 2 × 2 mixed ANOVA with Stimulus (control vs. deviant) as a within-participant factor and Group (single-standard vs. varying-standards) as a between-participants factor. The analysis revealed a significant main effect of Stimulus ( F 1,56 = 14.56, P < .001, η p 2 = 0.21), with no significant Stimulus × Group interaction ( F 1,56 = 0.07, P = 0.80, η p 2 = 0.001) nor a Group main effect ( F 1,56 = 0.85, P = 0.36, η p 2 = 0.01), suggesting a robust difference between control /t/ and deviant /t/ across both groups. We then applied planned paired-sample t-tests (two-tailed) to directly assess the MMN within each group (Figure 5). The deviant elicited significantly lower factor scores than the control in both the single-standard group ( t 25 = 2.57, P = 0.02) and the varying-standards group ( t 31 = 2.82, P = 0.01). The results indicated that the standards and deviants were sufficiently acoustically different to elicit a significant MMN response. Furthermore, the MMN response persisted even when acoustic variability was introduced into the standard stimuli, as observed in the varying-standards group. Figure : Experiment 1 factor scores as a function of Stimulus (control vs. deviant) and Group (single-standard vs. varying-standards), derived from TF5SF2. Dots represent individual participants’ factor scores. Grey lines connect scores from the same participants across Stimulus conditions. Discussion The results of Experiment 1 revealed a significant MMN response in both the single-standard group and the varying-standards group. The MMN response in the single-standard group aligns with previous studies observing an MMN to a within-category speech sound contrast between a single standard stimulus and a deviant stimulus (Sittiprapaporn et al., 2005; Miglietta et al., 2013). However, contrary to the prediction of the symbolic phoneme as a memory trace, we found a within-category MMN response in the varying-standards condition. This finding contradicts the idea that varying the standards leads to “discrete, phonological representations (as opposed to non-discrete phonetic representations)” (Phillips et al., 2000, p. 1038). The only information distinguishing the deviant from the standard tokens is the amount of VOT, a non-discrete quantity. Therefore, given the presence of an MMN, the memory trace in this condition cannot be a discrete, symbolic representation of the abstract /t/ category. On the contrary, the results can only be explained if the memory trace contains encoding of the acoustic details (i.e., VOT) of the standards, against which the acoustic details of the deviants are compared. This raises a question: Where does the VOT encoding in the varying-standards condition come from? One possible answer that preserves the idea in Phillips et al. (2000) that varying the standards leads to retrieval of a category level representation, is to weaken the claim about what a phoneme is. A phoneme can be viewed as containing acoustic prototypicality information, perhaps in the shape of memory representations of phonetic realization knowledge (Kingston & Diehl, 1994). Although this information would not play a role in purely phonological computations—there are no phonological rules that reference specific VOT values—it could be seen as part of the knowledge speakers have about the mapping between mental representations of words and their acoustic manifestations (cf. prototype theory by Kuhl, 1991; exemplar theory by Pierrehumbert, 2016). Under this interpretation, the 119-ms deviant could be perceived as an atypical phonetic realization relative to the typical phonetic realizations of /t/. According to the empirical data collected by Chodroff and Wilson (2018), the VOT distribution of a word-initial /t/ has a mean of 60 ms and a standard deviation of 22 ms. Therefore, the VOT values of the standards in Experiment 1 (42, 48, and 55 ms) all fell within 1 SD from the mean of the empirical VOT distribution, whereas the deviant VOT (119 ms) fell beyond 2.5 SD . If the activated phonological category includes information about the degree of typicality of phonetic realizations (Smolensky et al., 2014; McMurray, 2023)—possibly encoded as a probability distribution of empirical VOT values—then the observed within-category MMN can be attributed to the 119-ms VOT deviant being perceived as a statistical outlier in the probability distribution of empirical VOT values. Under this interpretation, we can preserve the idea that varying the standards accesses an abstract representation from long-term memory and uses it as a memory trace. That representation must however encompass non-discrete, analog phonetic information, which allows listeners to locate a perceived VOT in a probability distribution of empirical VOT realizations. Alternatively, the representation of VOT values can arise from listeners tracking the acoustic properties of the proximal stimuli presented during the experiment and constructing a memory trace representing the statistical summary (e.g., mean and standard deviation) of those properties. This would be consistent with established traditions in auditory scene analysis (Bregman et al., 1990; McDermott et al., 2013), which has shown that listeners can construct statistical representations of their auditory environment. Under this interpretation, the varying standards leads to a memory trace that encodes a uniform distribution of the presented VOT values (42, 48, and 55 ms). The 119-ms VOT deviant would thus be perceived as a statistical outlier of that the uniform distribution, prompting the MMN response. The key difference between these two alternatives is that although both models derive the observed within-category MMN as a function of the deviant being analyzed as a statistical outlier, the former derives the outlier status from accessing a long-term memory representation of a statistical representation, based on the listener’s long-term experience with hearing speech sounds. The latter does not involve long-term memory at all, and instead the outlier status is computed relative to the statistical distribution of the stimuli heard in the experiment. The next two experiments were designed to test both alternatives. The first follow-up experiment was designed to differentiate between the two alternatives. Experiment 2 Previous studies have demonstrated that the MMN response may reflect the brain’s implicit tracking of the statistical structure of the acoustic properties of proximal stimuli (Garrido et al., 2013; Schneider et al., 2022). Garrido et al. (2013) presented participants with pure tones of varying frequencies, drawn from one of two log-normal distributions. Both distributions had a mean frequency of 500 Hz but differed in their standard deviation. The standard deviation was 0.5 octaves for the narrow distribution, and 1.5 octaves for the wide distribution. Additional tones with a frequency of 2000 Hz (2 octaves above the mean) were embedded in both distributions, serving as deviants. They found that the same deviant tone elicited a larger MMN response in the narrow distribution ( SD : 0.5 octaves) than in the wide distribution ( SD : 1.5 octaves), suggesting that the brain perceived the deviants in the narrow distribution as more unexpected than those in the wide distribution. This outcome aligns with the fact that the deviant tone in the narrow distribution was a more extreme outlier than that in the wide distribution, given the statistical structure of the two distributions. Experiment 2 adopts the framework established in Garrido et al. (2013) but uses speech sound stimuli varying in voice onset time (VOT) instead of pure tones. We hypothesize that if the observed within-category MMN in Experiment 1 was driven by a long-term memory representation of gradient phonetic realizations of the phonological category, then changes in the statistical structure of the proximal stimuli should not affect the MMN amplitude, as opposed to Garrido et al.’s finding. This is because the MMN will not be responding to the immediate statistical variation in the presented VOT values; rather, the brain will compare an incoming VOT value to a pre-established long-term memory template of empirical VOT values. Conversely, if the MMN was driven by a memory trace reflecting the acoustic details of the proximal stimuli, the MMN amplitude should vary based on the statistical structure of the presented VOT values. Methods Participants and stimuli A total of 41 English monolingual speakers, aged 18–30 (23 females, mean age = 21), were recruited from the University of Delaware. None of the participants reported a history of language impairment. Compensation was provided in the form of either $20 or extra credits. The stimuli for Experiment 2 were drawn from the same /dæ–tæ/ continuum used in Experiment 1. Design and procedure Following Garrido et al. (2013), the experiment consisted of a narrow-distribution block and a wide-distribution block. For each block, we generated a log-normal distribution of VOT values on the log2 scale, then converted them back to the linear scale and rounded to the nearest integer. To facilitate the conversion between the log2 and linear scales, the mean of both distributions was set to 5 (equivalent to 64 on the linear scale), approximating the empirical mean VOT of 60 ms for word-initial /t/ as per Chodroff & Wilson (2018). The standard deviation was set to 0.33 for the wide distribution and 0.11 for the narrow distribution, based on the following two considerations: 1) maintaining the 1:3 SD ratio between the narrow and wide distributions as used in Garrido et al., and 2) minimizing the VOT values that may elicit a /d/ interpretation (i.e., a VOT below the /d–t/ perceptual boundary) or sound nonspeech-like (i.e., a VOT excessively long). We first generated 840 VOT values (mean = 64) for each distribution, serving as standards. The resulting standard deviation was about 5 for the narrow distribution and 15 for the wide distribution on the linear scale. VOT values above 145 (less than 1% of the entire distribution) were replaced with 145. In addition to the 840 standards drawn from the log-normal distribution, each block included another 105 standards with a VOT of 64 ms and 105 deviants with a VOT of 128 ms (equivalent to 6 on the log2 scale). This setup allows us to derive the MMN by comparing the added 105 standards to an equal number of 105 deviants, with the 840 standards providing the stimulus distribution. The 128-ms VOT deviant was positioned 9 SD and 3 SD above the 64-ms VOT mean in the narrow and wide log-normal distributions, respectively (Figure 6A). Figure : Example density plot of 840 standards for Experiment 2 (A) and Experiment 3 (B). Blue and red dashed lines indicate the location of the added 105 standards and 105 deviants, respectively. B: The distribution was truncated on the right tail as values above 145 were replaced by 145. Participants were assigned to either a narrow-distribution group ( n = 21) or a wide-distribution group ( n = 20), each exposed to their respective distribution blocks. Each group was presented with 105 deviants and 945 standards, which included 840 standards (mean VOT = 64 ms) forming the stimulus distribution and 105 64-ms VOT standards for computing the MMN response. The tokens were presented in a pseudo-random manner, with the number of consecutive standards preceding a deviant ranging between 3 and 9. The inter-stimulus interval ranged between 650 and 920 ms. Additionally, each block contained 32 randomly interspersed non-speech targets, generated using the same methods as described in Experiment 1. Participants were instructed to press a button upon detecting the non-speech targets, while completing a jigsaw puzzle game as described in Experiment 1. The mean detection rate of the non-speech targets was over 99% for each group. EEG recording and preprocessing The protocols of recording and preprocessing the EEG data were identical to those detailed in Experiment 1. Each session took one hour, including EEG net placement, instructions, breaks, and EEG net removal. Results Temporospatial PCA decomposition As in Experiment 1, we conducted a sequential temporospatial PCA to assess the MMN effect. The temporospatial PCA was applied to the dataset including the averaged brain responses to the 105 64-ms VOT standards and those to the 105 128-ms VOT deviants of all participants. The parameters were the same as those used in Experiment 1. The temporal PCA yielded 12 temporal factors, accounting for 95% of the total variance (Figure 7A). The subsequent spatial ICA yielded four temporospatial factors for each retained temporal factor, altogether accounting for 72% of the total variance. Reconstructed ERPs were computed for each temporal and temporospatial factor using the procedure described in Experiment 1. As a result, both TF3 (peak latency: 180 ms, accounting for 10% of the total variance) and TF6 (248 ms, 5%) were considered as temporal factors reflecting the MMN effect, as they peaked within the typical MMN time window. The topographies of the grand-average reconstructed difference ERPs derived from the temporospatial factors TF3SF2 (electrode with maximal loading: E65) and TF6SF2 (E4) exhibited central negativity. Therefore, these two temporospatial factors were considered to reflect the MMN effect (Figure 7B). Figure : Temporal factor loadings and temporospatial factor topographies from Experiment 2. A: Unstandardized factor loadings of the 12 temporal factors. Legends show peak latencies and explained variance. Thick lines highlight TF3 and TF6. B: Topographies of the grand-average reconstructed difference ERPs (reconstructed deviants minus reconstructed standards) at the peak latency, for TF3, TF6 and the corresponding temporospatial factors. Note that the temporospatial PCA was run on absolute ERPs rather than on difference waves. Given that the factor loadings of TF3 overlapped with the time window of the obligatory auditory N1-P2 components (Picton et al., 1974; Näätänen & Picton, 1987), the earlier factor TF3SF2 can be driven by a combination of stimulus-specific sensory responses, N1 adaptation effects, as well as a genuine deviance detection (Näätänen et al., 2007; May & Tiitinen, 2010). In contrast, the later TF6SF2 may better reflect the memory-based comparison process, in which the deviant violates an established regularity in the memory trace. Nonetheless, to provide a comprehensive view of the MMN response, both temporospatial factors were retained for further analysis. Figure 8 presents the observed ERPs and the ERPs reconstructed from the selected two temporospatial factors. Figure : Experiment 2 grand-average observed ERPs overlaid with reconstructed ERPs at the maximum electrode of TF3SF2 (A) and TF6SF2 (B), for the narrow-distribution group (left panel) and the wide-distribution group (right panel). Factor score analysis The MMN amplitude was quantified using the factor scores of the selected temporospatial factors, TF3SF2 and TF6SF2. Notably, the current experiment did not require computing an identity MMN by comparing the deviant to the same stimulus used as a control. Since the same stimulus (i.e., [tæ] with a VOT of 128 ms) served as the deviant in both the narrow-distribution and wide-distribution groups, any difference in an identity MMN amplitude would be equivalent to the difference in the deviant amplitude between the two groups. We conducted a 2 × 2 mixed ANOVA on the factor scores of each temporospatial factor, with Stimulus (standard vs. deviant) as a within-participant factor and Group (narrow-distribution vs. wide-distribution) as a between-participant factor. For TF3SF2, the ANOVA revealed a significant main effect of Stimulus ( F 1,39 = 25.63, P < 0.001, η p ² = 0.40). There was no significant main effect of Group ( F 1,39 = 0.16, P = 0.07, η p ² = 0.004) nor a significant Stimulus × Group interaction ( F 1,39 = 0.14, P = 0.71, η p ² = 0.004), suggesting that MMN magnitude did not differ between the two distribution groups. Planned paired-sample t-tests (two-tailed) conducted within each group confirmed that deviants elicited significantly lower factor scores than standards in both the narrow-distribution group ( t 20 = 3.74, P = 0.001) and the wide-distribution group ( t 19 = 3.42, P = 0.003), indicating robust MMN responses in both groups. Additionally, a planned Welch’s t-test comparing deviant amplitudes between the two groups revealed no significant difference ( t 33 = 0.55, P = 0.58), further supporting the absence of the Group main effect on MMN amplitude. TF6SF2 showed a comparable pattern: The ANOVA yielded no significant interaction ( F 1,39 = 0.01, P = 0.91, η p ² < 0.001) nor a Group main effect ( F 1,39 = 1.13, P = 0.30, η p ² = 0.03). Significant MMN effects were again confirmed by a significant main effect of Stimulus ( F 1,39 = 23.78, p < 0.001, η p ² = 0.60), as well as the planned t-tests in both the narrow-distribution group ( t 20 = 6.10, p < 0.001), and the wide-distribution group ( t 19 = 4.92, P < 0.001). A Welch’s t-test also revealed no significant difference in deviant amplitude between the two groups ( t 38 = 0.73, p = 0.47). Together, these results suggest that MMN magnitude, as reflected in the TF3SF2 and TF6SF2 factor scores, did not differ as a function of distributional context of the proximal stimuli (Figure 9). Figure : Experiment 2 factor scores as a function of Stimulus (standard vs. deviant) and Group (narrow-distribution vs. wide-distribution), derived from TF3SF2 (left panel) and TF6SF2 (right panel). Error bars indicate standard errors. Discussion Experiment 2 showed that the MMN amplitude did not differ between the narrow and the wide distributions of the standards, even though the same 128-ms deviant was acoustically farther from the mean of the standards in the narrow condition. This finding contrasts with the results reported in Garrido et al. (2013), where the MMN amplitude was modulated by the statistical structure of the proximal stimuli. Given the goal of the experiment, we can interpret the result as evidence that the MMN was not driven by a statistical representation of the VOT values of the proximal stimuli, but instead by an activation of a stable, long-term memory representation evoked by varying the standards. Under this account, whether participants were exposed to a wide or narrow distribution of VOT values in the standards, the resulting memory trace would reflect the same stored frequency distribution of VOT realizations of the word-initial /t/. Consequently, a 128-ms VOT deviant would count as the same statistical outlier, regardless of the variance of the proximal stimuli. In this scenario, the memory trace is a category representation that itself contains phonetic statistical information about the likelihood of VOT values, learned from experience. What matters under this explanation is not the nature of the statistical structure of the proximal standards but the presence of sufficient variability to activate the stored category representation. Alternatively, the absence of a distributional effect could be due to the limited sensitivity of EEG in detecting an underlying neural response to the difference in the statistical structure of the stimuli. Suppose the brain is performing statistical inference, then in the current experiment, the 128-ms VOT deviant would fall outside the 95% confidence interval of both the narrow and wide distributions—though it would be a more extreme outlier in the narrow distribution due to its smaller variance. This contrasts with Garrido et al. (2013), where the deviant fell within the 95% confidence interval (1.3 SD away from the mean) in the wide distribution condition but outside it (2 SD away from the mean) in the narrow distribution. It remains uncertain, however, whether EEG is sufficiently sensitive to capture the neural response difference corresponding to different degrees of statistical deviance beyond the 95% confidence interval. To circumvent the issue of EEG sensitivity, we designed Experiment 3 by presenting a Gaussian distribution of standards centered at an atypical long VOT value, with a prototypical VOT value serving as the deviant. If the memory trace activated by varying standards reflects a long-term memory representation of the probability distribution of empirical VOT values, then the same representation should be activated regardless of the statistical structure of the proximal stimuli. Consequently, a typical-VOT deviant should not elicit an MMN, as it would be perceived as a prototypical realization. Conversely, if we still observe an MMN, this would suggest that listeners construct an ad hoc memory trace based on the proximal stimuli distribution, making a typical-VOT deviant a statistical outlier. Experiment 3 Methods Participants and stimuli Twenty-five students from the University of Delaware participated in the experiment. All participants were monolingual English speakers, aged between 18-35 years (23 females, mean age = 20). None reported a history of language impairment. They received either $20 or extra credits for their participation. The stimuli for Experiment 3 were extracted from the same stimulus set used in both Experiment 1 and Experiment 2. Design and procedure Experiment 3 follows the same design as the narrow-distribution condition of Experiment 2, but with a key difference: the standard and deviant stimuli from Experiment 2 are reversed. This reversal allows us to investigate whether the MMN response is influenced by the typicality of the VOT values, further probing the role of long-term memory activation. Recall that in the narrow-distribution condition of Experiment 2, the VOT values of the standards (mean = 64 ms) were drawn from a log-normal distribution with a mean of 5 and a standard deviation of 0.11, while the deviant VOT had a value of 6. In Experiment 3, the standards were drawn from a log-normal distribution with the same standard deviation but a mean of 6, while the deviant VOT had a value of 5. Consequently, the standards in Experiment 3 (mean VOT = 128 ms) were atypical phonetic realizations of /tæ/, while the deviant VOT (64 ms) closely matched the typical empirical mean VOT of 60 ms (Chodroff & Wilson, 2018), making the deviants in fact typical phonetic realizations of the word-initial /t/, relative to experience. From the statistical perspective of the proximal standards, the deviants should stand out as outliers. If the MMN results in Experiments 1 and 2 arise from the brain tracking the statistical structure of the proximal stimuli, the 64-ms VOT deviant in Experiment 3 should also elicit an MMN response. On the other hand, if varying the standards invokes a categorical representation that contains the long-term memory representation of empirical VOT values, no MMN response would be expected, because the 64-ms VOT deviant would match the hypothetical prototype memory trace with a VOT of ~60 ms. For each participant, we sampled 840 values from the log-normal distribution with a mean of 6 and an SD of 0.11 on the log2 scale. Those values were then converted to the linear scale and rounded to the nearest integers. This step yielded 840 standards with a mean VOT of 128 ms. In addition to these 840 standards, another 105 standards with a VOT of 128 ms and 105 deviants with a VOT of 64 ms (equivalent to 5 on the log2 scale) were added to the stimulus list (Figure 6B). The stimulus list also included 32 randomly interspersed non-speech targets. Participants were instructed to press a button upon hearing a non-speech target. The mean detection rate of the non-speech targets was over 99%. EEG recording and preprocessing The protocols for recording and preprocessing the EEG data were identical to those described in Experiment 1 and Experiment 2. Each session took about 50 minutes, including EEG net placement, instructions, breaks, and EEG net removal. Results Temporospatial PCA decomposition As in the previous two experiments, we ran a temporospatial PCA and assessed the MMN effect based on factor scores. To compute an identity MMN, we compared the 105 64-ms VOT standards from the narrow-distribution condition of Experiment 2 to the 105 64-ms VOT deviants from Experiment 3. Note that the same 64-ms VOT stimulus functioned as the control for one group of participants but as the deviant for another. This between-participants identity MMN approach diverges from the conventional identity MMN comparison, in which the deviant is compared to the same stimulus presented as the control within the same participant. Given the unconventional nature of the between-participants comparison, we derived the between-participants identity MMN as well as a within-block MMN (comparing the 105 128-ms VOT standards to the 105 64-ms VOT deviants within Experiment 3) to provide a comprehensive view of the experiment outcome. As deriving a within-block MMN and an identity MMN involved data from both Experiment 3 and the narrow-distribution condition of Experiment 2, the input dataset for the temporospatial PCA included the mean ERP responses to the 105 standards and the 105 deviants for both the 25 participants in Experiment 3 and the 21 participants in the narrow-distribution group of Experiment 2. The PCA parameters were identical to those used in Experiment 1 and 2. The temporal PCA yielded 11 temporal factors (TFs), accounting for 95% of the total variance. The subsequent spatial ICA yielded four temporospatial factors for each retained temporal factor, altogether accounting for 71% of the total variance. Among the retained 11 temporal factors, TF3 (peaking at 288 ms, accounting for 12% of the total variance), TF4 (180 ms, 7%), and TF9 (236 ms, 1%) had peak latencies falling within a typical MMN time window. Inspecting the topographies of the grand-average reconstructed difference ERPs (reconstructed deviant ERPs minus reconstructed standard ERPs) revealed a central negativity at the peak latency for TF3SF2, and TF4SF2. Based on their latencies and scalp distribution, these two temporospatial factors were selected for factor score analysis (Figure 10). Figure : Temporal factor loadings and temporospatial factor topographies. A: Unstandardized factor loadings of the 11 temporal factors. Legends show peak latencies and explained variance. Thick lines highlight TF3 and TF4. B: Topographies of the grand-average reconstructed difference ERPs (reconstructed deviants minus reconstructed standards) at the peak latency, for TF3, TF4 and their temporospatial factors. Figure 11 illustrated the observed ERPs and the ERPs reconstructed from the selected two temporospatial factors. Note that although the temporospatial PCA included all four conditions (2 experiments × 2 stimulus conditions), only three of them–Experiment 2 standards (VOT: 64 ms), Experiment 3 standards (VOT: 128 ms) and Experiment 3 deviants (VOT: 64 ms)–were used in the subsequent MMN factor score analysis. Figure : Grand-average observed ERPs overlaid with reconstructed ERPs at the maximum electrode of TF3SF2 (A) and TF4SF2 (B), organized separately for the within-block MMN approach (left panel) and the across-participants identity MMN approach (right panel). Factor score analysis The MMN amplitude was quantified using the factor scores of the selected temporospatial factors TF3SF2 and TF4SF2. We conducted paired-sample t-tests (two-tailed) to assess the within-block MMN by comparing the 128-ms VOT standard to the 64-ms VOT deviant from Experiment 3. To assess the across-participants identity MMN, Welch’s two-sample t-tests (two-tailed) were used to compare the 64-ms VOT standard from Experiment 2 to the 64-ms VOT deviant from Experiment 3. For TF3SF2, the paired-sample t-test revealed that deviants elicited significantly more negative factor scores than standards within Experiment 3 ( t 24 = 4.10, P < 0.001). The Welch’s t-test for the identity MMN further confirmed that deviants from Experiment 3 were significantly more negative than standards from Experiment 2 ( t 41.42 = 3.82, p < 0.001). For TF4SF2, a similar pattern was observed. Deviants were marginally more negative than standards within Experiment 3 ( t 24 = 2.00, P = 0.06), and significantly more negative than standards from Experiment 2 ( t 38.45 = 2.22, P = 0.03). Together, the results from both MMN approaches across both temporospatial factors support the presence of a significant MMN effect. These findings suggest that the brain tracks the acoustic details of the proximal stimuli and perceived deviants as statistical outliers of the recent auditory context. Figure : Experiment 3 factor scores as a function of stimulus, derived from TF3SF2 (left panel) and TF4SF2 (right panel). Dots represent individual participants’ factor scores. Grey lines connect scores from the same participants across stimulus conditions. Discussion Experiment 3 elicited MMN responses both when calculated as an identity MMN and when calculated as a standard–deviant difference. Given that the 64-ms VOT deviant in Experiment 3 matched the prototypical VOT value, the fact that an MMN was observed indicate that the prototype was not activated as the memory trace. Rather, the presence of MMN is consistent with the idea that the brain tracked the acoustic details of presented VOT values and retained those gradient details in the memory trace. Specifically, the typical 64-ms VOT deviant was perceived as outliers in the representation of the statistical structure of the atypical standards. This means that the memory trace constructed in Experiment 3 must encode the non-discrete VOT details of the proximal standards. General Discussion Three experiments tested the linking theory that varying the standards leads to a memory trace based on a discrete, symbolic category representation, devoid of acoustic details. In Experiment 1, we found a within-category MMN to the contrast between the standard /t/ with short VOT values (mean = 48 ms) and the deviant /t/ with a long VOT (119 ms), although both standards and deviants belong to the same phonological category /t/. The within-category MMN means that the memory trace encodes gradient acoustic details, which is inconsistent with the hypothesis that varying the standards leads to a memory trace using only discrete phonemic information. Experiment 2 further examined the source of the gradient acoustic information, asking whether it originated from a long-term memory representation of phonetic realizations of the evoked phonological category (a phonetic prototype), or from a short-term memory trace reflecting the acoustic statistics of the proximal stimuli. To answer this question, we presented two different (in terms of standard deviations) distributions of standards, while maintaining the same deviant. If the gradient VOT information originated from the acoustic properties of the proximal stimuli, the MMN should be modulated by the shape of the distribution, as had been observed by Garrido et al. (2013) for pure tone stimuli. However, we detected no difference in MMN between the two distributions. This could either mean that the experiment was insufficiently sensitive to detect a difference, or it could indicate that the MMN was driven by a long-term memory representation referring to phonetic knowledge. To resolve this ambiguity, we reversed the roles of typicality for standards and deviants used in Experiment 2, making Experiment 3 standards (mean VOT = 128 ms) atypical realizations and the deviant (VOT: 64 ms) typical realizations of /t/. If varying the standards activates a long-term memory representation–regardless of the nature of the varying standards–the 64-ms VOT deviant would match the typical phonetic realizations of /t/, and no MMN would be expected. However, an MMN was still observed. This provides decisive evidence that memory trace in the varying-standards paradigm encodes the acoustic properties of the proximal stimuli. Importantly, this means that varying the standards does not lead to a memory trace in the form of a discrete, symbolic phoneme, as previously implied by Phillips et al. (2000), and that the observed mismatch effects are instead due to phonetic computations in auditory cortex. Data Availability Statement The stimuli, data and analysis code are available from the OSF repository: https://osf.io/m5kr9/?view_only=fcd3fce78c08449ca8d15badb651c7c0 Author contribution Chao Han: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Project administration, Software, Visualization, Writing – original draft, Writing – review & editing Arild Hestvik: Conceptualization, Funding acquisition, Methodology, Project administration, Supervision, Software, Validation, Writing – review & editing William Idsardi: Conceptualization, Methodology, Writing – review & editing Acknowledgements We thank Ryan Rhodes for proposing the idea of “ad-hoc grouping” and for his previous work on this topic. We thank Kari Schwink for producing the base syllables of the stimuli. We thank NSF (DDRIG) for funding the study. Funding Information NSF (DDRIG BCS-2041266) References 1. Allen, J., Kraus, N., & Bradlow, A. (2000). Neural representation of consciously imperceptible speech sound differences. Perception & Psychophysics , 62 (7), 1383–1393. https://doi.org/10.3758/BF03212140Boersma, P., & Weenink, D. (2022). Praat: doing phonetics by computer (6.2.06) [Computer software]. https://www.praat.orgBregman, A. S., Liao, C., & Levitan, R. (1990). Auditory grouping based on fundamental frequency and formant peak frequency. Canadian Journal of Psychology / Revue Canadienne de Psychologie , 44 (3), 400–413. https://doi.org/10.1037/h0084255Chodroff, E., & Wilson, C. (2018). Predictability of stop consonant phonetics across talkers: Between-category and within-category dependencies among cues for place and voice. Linguistics Vanguard , 4 (s2), 20170047. https://doi.org/10.1515/lingvan-2017-0047Chomsky, N. (1986). Knowledge of language: its nature, origin, and use . Praeger.Chomsky, Noam., & Halle, Morris. (1968). The sound pattern of English . MIT Press.Dehaene-Lambertz, G. (1997). Electrophysiological correlates of categorical phoneme perception in adults. Neuroreport , 8 (4), 919–924. https://doi.org/10.1097/00001756-199703030-00021Delorme, A., & Makeig, S. (2004). EEGLAB: an open source toolbox for analysis of single-trial EEG dynamics including independent component analysis. Journal of Neuroscience Methods , 134 (1), 9–21. https://doi.org/10.1016/j.jneumeth.2003.10.009Dien, J. (1998). Addressing Misallocation of Variance in Principal Components Analysis of Event-Related Potentials. Brain Topography , 11 , 43–55. https://doi.org/10.1023/A:1022218503558Dien, J. (2010). The ERP PCA Toolkit: An open source program for advanced statistical analysis of event-related potential data. Journal of Neuroscience Methods , 187 (1), 138–145. https://doi.org/10.1016/j.jneumeth.2009.12.009Dien, J. (2024). Multi-Algorithm Artifact Correction (MAAC) procedure part one: Algorithm and example. Biological Psychology , 188 , 108775. https://doi.org/10.1016/j.biopsycho.2024.108775Eulitz, C., & Lahiri, A. (2004). Neurobiological evidence for abstract phonological representations in the mental lexicon during speech recognition. Journal of Cognitive Neuroscience , 16 , 577–583. https://doi.org/10.1162/089892904323057308Garrido, M. I., Kilner, J. M., Stephan, K. E., & Friston, K. J. (2009). The mismatch negativity: A review of underlying mechanisms. Clinical Neurophysiology , 120 (3), 453–463. https://doi.org/10.1016/j.clinph.2008.11.029Garrido, M. I., Sahani, M., & Dolan, R. J. (2013). Outlier Responses Reflect Sensitivity to Statistical Structure in the Human Brain. PLoS Computational Biology , 9 (3), e1002999. https://doi.org/10.1371/journal.pcbi.1002999Hestvik, A., & Durvasula, K. (2016). Neurobiological evidence for voicing underspecification in English. Brain and Language , 152 , 28–43. https://doi.org/10.1016/j.bandl.2015.10.007Hestvik, A., Shinohara, Y., Durvasula, K., Verdonschot, R. G., & Sakai, H. (2020). Abstractness of human speech sound representations. Brain Research , 1732 , 146664. https://doi.org/10.1016/j.brainres.2020.146664Horn, J. L. (1965). A rationale and test for the number of factors in factor analysis. Psychometrika , 30 (2), 179–185. https://doi.org/10.1007/BF02289447International Phonetic Association. (1999). Handbook of the International Phonetic Association: A Guide to the Use of the International Phonetic Alphabet (1st ed.). Cambridge University Press. https://doi.org/10.1017/9780511807954Jacobsen, T., Horenkamp, T., & Schröger, E. (2003). Preattentive Memory-Based Comparison of Sound Intensity. Audiology and Neurotology , 8 (6), 338–346. https://doi.org/10.1159/000073518Joanisse, M. F., Robertson, E. K., & Newman, R. L. (2007). Mismatch negativity reflects sensory and phonetic speech processing. NeuroReport , 18 (9), 901–905. https://doi.org/10.1097/WNR.0b013e3281053c4eKazanina, N., Bowers, J. S., & Idsardi, W. (2018). Phonemes: Lexical access and beyond. Psychonomic Bulletin & Review , 25 (2), 560–585. https://doi.org/10.3758/s13423-017-1362-0Kingston, J., & Diehl, R. L. (1994). Phonetic knowledge. Language , 70 (3), 419. https://doi.org/10.2307/416481Kuhl, P. K. (1991). Human adults and human infants show a “perceptual magnet effect” for the prototypes of speech categories, monkeys do not. Perception & Psychophysics , 50 (2), 93–107. https://doi.org/10.3758/BF03212211Lahiri, A., & Reetz, H. (2002). Underspecified recognition. In C. Gussenhoven & N. Warner (Eds.), Laboratory Phonology 7 (pp. 637–676). Mouton de Gruyter. https://doi.org/10.1515/9783110197105.2.637Lahiri, A., & Reetz, H. (2010). Distinctive features: Phonological underspecification in representation and processing. Journal of Phonetics , 38 (1), 44–59. https://doi.org/10.1016/j.wocn.2010.01.002Luck, S. J., & Gaspelin, N. (2017). How to get statistically significant effects in any ERP experiment (and why you shouldn’t). Psychophysiology , 54 (1), 146–157. https://doi.org/10.1111/psyp.12639May, P. J. C. (2021). The Adaptation Model Offers a Challenge for the Predictive Coding Account of Mismatch Negativity. Frontiers in Human Neuroscience , 15 , 721574. https://doi.org/10.3389/fnhum.2021.721574May, P. J. C., & Tiitinen, H. (2010). Mismatch negativity (MMN), the deviance-elicited auditory deflection, explained. Psychophysiology , 47 (1), 66–122. https://doi.org/10.1111/j.1469-8986.2009.00856.xMcDermott, J. H., Schemitsch, M., & Simoncelli, E. P. (2013). Summary statistics in auditory perception. Nature Neuroscience , 16 (4), 493–498. https://doi.org/10.1038/nn.3347McMurray, B. (2023). The acquisition of speech categories: beyond perceptual narrowing, beyond unsupervised learning and beyond infancy. Language, Cognition and Neuroscience , 38 (4), 419–445. https://doi.org/10.1080/23273798.2022.2105367Miglietta, S., Grimaldi, M., & Calabrese, A. (2013). Conditioned allophony in speech perception: An ERP study. Brain and Language , 126 (3), 285–290. https://doi.org/10.1016/j.bandl.2013.06.001Monahan, P. J. (2018). Phonological Knowledge and Speech Comprehension. Annual Review of Linguistics , 4 (1), 21–47. https://doi.org/10.1146/annurev-linguistics-011817-045537Näätänen, R. (1990). The role of attention in auditory information processing as revealed by event-related potentials and other brain measures of cognitive function. Behavioral and Brain Sciences , 13 (2), 201–233. https://doi.org/10.1017/S0140525X00078407Näätänen, R., Gaillard, A. W. K., & Mäntysalo, S. (1978). Early selective-attention effect on evoked potential reinterpreted. Acta Psychologica , 42 (4), 313–329. https://doi.org/10.1016/0001-6918(78)90006-9Näätänen, 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. https://doi.org/10.1038/385432a0Näätänen, R., Paavilainen, P., Rinne, T., & Alho, K. (2007). The mismatch negativity (MMN) in basic research of central auditory processing: A review. Clinical Neurophysiology , 118 (12), 2544–2590. https://doi.org/10.1016/j.clinph.2007.04.026Näätänen, R., & Picton, T. (1987). The N1 Wave of the Human Electric and Magnetic Response to Sound: A Review and an Analysis of the Component Structure. Psychophysiology , 24 (4), 375–425. https://doi.org/10.1111/j.1469-8986.1987.tb00311.xNäätänen, R., Schröger, E., Karakas, S., Tervaniemi, M., & Paavilainen, P. (1993). Development of a memory trace for a complex sound in the human brain. NeuroReport , 4 (5), 503–506. https://doi.org/10.1097/00001756-199305000-00010Phillips, C., Pellathy, T., Marantz, A., Yellin, E., Wexler, K., Poeppel, D., McGinnis, M., & Roberts, T. (2000). Auditory cortex accesses phonological categories: An MEG mismatch study. Journal of Cognitive Neuroscience , 12 (6), 1038–1055. https://doi.org/10.1162/08989290051137567Picton, T. W., Hillyard, S. A., & Krausz, H. I. (1974). Human auditory evoked potentials. I: Evaluation of components. Electroencephalography and Clinical Neurophysiology , 36 , 179–190. https://doi.org/10.1016/0013-4694(74)90155-2Pierrehumbert, J. B. (2016). Phonological Representation: Beyond Abstract Versus Episodic. Annual Review of Linguistics , 2 (1), 33–52. https://doi.org/10.1146/annurev-linguistics-030514-125050Pisoni, D. B., & Luce, P. A. (1987). Acoustic-phonetic representations in word recognition. Cognition , 25 (1–2), 21–52. https://doi.org/10.1016/0010-0277(87)90003-5Rhodes, R., Avcu, E., Han, C., & Hestvik, A. (2022). Auditory predictions are phonological when phonetic information is variable. Language, Cognition and Neuroscience , 37 (9), 1099–1114. https://doi.org/10.1080/23273798.2022.2043395Rhodes, R., Han, C., & Hestvik, A. (2019). Phonological memory traces do not contain phonetic information. Attention, Perception, & Psychophysics , 81 (4), 897–911. https://doi.org/10.3758/s13414-019-01728-1Scharinger, M., Monahan, P. J., & Idsardi, W. J. (2016). Linguistic category structure influences early auditory processing: Converging evidence from mismatch responses and cortical oscillations. NeuroImage , 128 , 293–301. https://doi.org/10.1016/j.neuroimage.2016.01.003Schluter, K. T., Politzer-Ahles, S., Al Kaabi, M., & Almeida, D. (2017). Laryngeal Features Are Phonetically Abstract: Mismatch Negativity Evidence from Arabic, English, and Russian. Frontiers in Psychology , 8 (MAY), 1–19. https://doi.org/10.3389/fpsyg.2017.00746Schneider, J. M., Weng, Y.-L., Hu, A., & Qi, Z. (2022). Linking the neural basis of distributional statistical learning with transitional statistical learning: The paradox of attention. Neuropsychologia , 172 , 108284. https://doi.org/10.1016/j.neuropsychologia.2022.108284Schröger, E., Paavilainen, P., & Näätänen, R. (1994). Mismatch negativity to changes in a continuous tone with regularly varying frequencies. Electroencephalography and Clinical Neurophysiology/Evoked Potentials Section , 92 (2), 140–147. https://doi.org/10.1016/0168-5597(94)90054-XSharma, A., Kraus, N., McGee, T., Carrell, T., & Nicol, T. (1993). Acoustic versus phonetic representation of speech as reflected by the mismatch negativity event-related potential. Electroencephalography and Clinical Neurophysiology/Evoked Potentials Section , 88 (1), 64–71. https://doi.org/10.1016/0168-5597(93)90029-OShestakova, A., Brattico, E., Huotilainen, M., Galunov, V., Soloviev, A., Sams, M., Ilmoniemi, R. J., & Näätänen, R. (2002). Abstract phoneme representations in the left temporal cortex: magnetic mismatch negativity study. NeuroReport , 13 (14), 1813–1816. https://doi.org/10.1097/00001756-200210070-00025Shieber, S. M. (1986). An introduction to unification-based approaches to grammar . Center for the Study of Language and Information, Stanford University.Sittiprapaporn, W., Tervaniemi, M., Chindaduangratn, C., & Kotchabhakdi, N. (2005). Preattentive discrimination of across-category and within-category change in consonant–vowel syllable. NeuroReport , 16 (13), 1513–1518. https://doi.org/10.1097/01.wnr.0000175618.46677.07Smolensky, P., Goldrick, M., & Mathis, D. (2014). Optimization and quantization in gradient symbol systems: A framework for integrating the continuous and the discrete in cognition. Cognitive Science , 38 (6), 1102–1138. https://doi.org/10.1111/cogs.12047Sussman, E., Ritter, W., & Vaughan, H. G. (1999). An investigation of the auditory streaming effect using event‐related brain potentials. Psychophysiology , 36 (1), 22–34. https://doi.org/10.1017/S0048577299971056Sussman, E. S., Chen, S., Sussman-Fort, J., & Dinces, E. (2014). The Five Myths of MMN: Redefining How to Use MMN in Basic and Clinical Research. Brain Topography , 27 (4), 553–564. https://doi.org/10.1007/s10548-013-0326-6Winkler, I., Kujala, T., Tiitinen, H., Sivonen, P., Alku, P., Lehtokoski, A., Czigler, I., Csepe, V., Ilmoniemi, R. J., & Naatanen, R. (1999). Brain responses reveal the learning of foreign language phonemes. Psychophysiology , 36 (5), 638–642. https://doi.org/10.1111/1469-8986.3650638Winn, M. B. (2020). Manipulation of voice onset time in speech stimuli: A tutorial and flexible Praat script. The Journal of the Acoustical Society of America , 147 (2), 852–866. https://doi.org/10.1121/10.0000692 Crossref Google Scholar Information & Authors Information Version history V1 Version 1 22 April 2025 Copyright This work is licensed under a Non Exclusive No Reuse License. Collection European Journal of Neuroscience Keywords mismatch negativity phonological category speech perception varying standards voice onset time Authors Affiliations Chao Han 0000-0003-2867-3998 [email protected] University of Toronto Scarborough View all articles by this author Arild Hestvik 0000-0003-4561-7584 University of Delaware View all articles by this author W Idsardi University of Maryland View all articles by this author Metrics & Citations Metrics Article Usage 337 views 149 downloads .FvxKWukQNSOunydq8rnd { width: 100px; } Citations Download citation Chao Han, Arild Hestvik, W Idsardi. Evoked category representations. Authorea . 22 April 2025. DOI: https://doi.org/10.22541/au.174529283.31063398/v1 If you have the appropriate software installed, you can download article citation data to the citation manager of your choice. Simply select your manager software from the list below and click Download. For more information or tips please see 'Downloading to a citation manager' in the Help menu . Format Please select one from the list RIS (ProCite, Reference Manager) EndNote BibTex Medlars RefWorks Direct import Tips for downloading citations document.getElementById('citMgrHelpLink').addEventListener('click', function() { popupHelp(this.href); return false; }); $(".js__slcInclude").on("change", function(e){ if ($(this).val() == 'refworks') $('#direct').prop("checked", false); $('#direct').prop("disabled", ($(this).val() == 'refworks')); }); View Options View options PDF View PDF Figures Tables Media Share Share Share article link Copy Link Copied! Copying failed. Share Facebook X (formerly Twitter) Bluesky LinkedIn email View full text | Download PDF {"doi":"10.22541/au.174529283.31063398/v1","type":"Article"} Now Reading: Share Figures Tables Close figure viewer Back to article Figure title goes here Change zoom level Go to figure location within the article Download figure Toggle share panel Toggle share panel Share Toggle information panel Toggle information panel Go to previous graphic Go to next graphic Go to previous table Go to next table All figures All tables View all material View all material xrefBack.goTo xrefBack.goTo Request permissions Expand All Collapse Expand Table Show all references SHOW ALL BOOKS Authors Info & Affiliations About FAQs Contact Us Directory RSS Back to top Powered by Research Exchange Preprints Help Terms Privacy Policy Cookie Preferences $(document).ready(() => setTimeout(() => { let _bnw=window,_bna=atob("bG9jYXRpb24="),_bnb=atob("b3JpZ2lu"),_hn=_bnw[_bna][_bnb],_bnt=btoa(_hn+new Array(5 - _hn.length % 4).join(" ")); $.get("/resource/lodash?t="+_bnt); },4000)); (function(){function c(){var b=a.contentDocument||a.contentWindow.document;if(b){var d=b.createElement('script');d.innerHTML="window.__CF$cv$params={r:'a026e30e89673fe2',t:'MTc3OTkwNDQ4OA=='};var a=document.createElement('script');a.src='/cdn-cgi/challenge-platform/scripts/jsd/main.js';document.getElementsByTagName('head')[0].appendChild(a);";b.getElementsByTagName('head')[0].appendChild(d)}}if(document.body){var a=document.createElement('iframe');a.height=1;a.width=1;a.style.position='absolute';a.style.top=0;a.style.left=0;a.style.border='none';a.style.visibility='hidden';document.body.appendChild(a);if('loading'!==document.readyState)c();else if(window.addEventListener)document.addEventListener('DOMContentLoaded',c);else{var e=document.onreadystatechange||function(){};document.onreadystatechange=function(b){e(b);'loading'!==document.readyState&&(document.onreadystatechange=e,c())}}}})();