Semantic processing of foreign-accented speech

In our globalised world it is common for language users to encounter speakers of their language who have a foreign accent. This is particularly the case with languages such as English, which are commonly used as a lingua franca among speakers of diverse linguistic backgrounds. Worldwide, 64% of English speakers are non-native speakers of English [2], and thus have a high likelihood of speaking with a foreign accent [3]. Even in predominantly English-speaking countries, the percentage of immigrant, and therefore potentially foreign-accented, speakers can be substantial. In Sydney, Australia, for example, 42% of the population are born outside Australia [4].

An understanding of how foreign-accented speech is processed is therefore essential for an ecologically valid picture of language processing. The impact of a foreign accent on some aspects of speech processing has been examined in detail (particularly phonological processing), while other aspects remain little studied. One such area is semantic processing, which involves the online process of retrieving and integrating the meaning of lexical items.

A small number of studies have used event-related potentials (ERPs) and the semantic-violation paradigm to explore the semantic processing of foreign-accented speech [1,5,6]. ERPs enable investigation of listeners’ brain responses to language input with excellent temporal precision via the measurement of scalp-recorded electrical activity. The resulting ERP component can be described in relation to the following characteristics: polarity, amplitude, latency, onset latency and topographical distribution. ERP components can be correlated to specific cognitive behaviours associated with the processing of the stimulus [7]. In ERP studies of semantic processing, a semantic-violation paradigm is typically employed in which participants read or listen to words that are either semantically congruent or incongruent with the preceding sentential/discourse context while their brain activity is recorded (e.g. ‘cupcake’ in The lady bites the cupcake vs. ‘suitcase’ in The lady bites the *suitcase). Generally, semantically incongruent sentences elicit a larger negativity than semantically congruent sentences, peaking around 400 ms post-stimulus-onset and with a centro-parietal maximum ([8], see [9] for a review). This negativity is referred to as the N400 component and the difference in N400 amplitude between two conditions is termed an N400 effect. This effect is generally interpreted as being an index of semantic-memory retrieval or semantic integration difficulty [9,10].

There appears to be a consensus that the N400 effects elicited by semantic violations in native and foreign-accented speech differ. The nature of the difference has varied across studies but has primarily involved the amplitude and topography of the N400. Hanuliková and colleagues [1] found that the N400 had a broader topographical distribution in response to foreign-accented than native Dutch, but did not differ in amplitude. In partial contrast, Romero-Rivas, Martin and Costa [5] found that the N400 effect for foreign-accented Spanish was both more broadly distributed and of greater amplitude than that observed for native Spanish. Finally, Grey and Van Hell [6] found that the N400 effect for foreign-accented speech had a later onset latency than that for native English, but were unable to directly compare the two N400s in amplitude and topography as they occurred in different time windows. (Note that due to this timing difference the N400 for foreign-accented speech could alternatively be interpreted as a functionally distinct late negativity).

Aside from these studies of N400 effects elicited by semantic violations, there has also been some examination of N400 effects elicited by foreign-accented sentences varying in cloze probability and in semantically congruent sentences varying in accentedness. Regarding cloze probability, similar N400 effects were found when comparing foreign-accented sentences with the semantic best completion to sentences with both semantically related and unrelated completions [11]. This differed from the pattern for native speech, which saw a gradient effect of semantic relatedness on N400 amplitude. Regarding sentence accentedness, an N400 effect was also found when comparing semantically congruent sentences produced in a native accent to semantically congruent sentences produced with a foreign accent [12]. However, as these studies examined N400 effects elicited by phenomena other than semantic violations, they are not directly comparable to the studies outlined above, or the current study.

Previous results thus consistently demonstrate a difference in the N400 effects elicited by semantic violations in native and foreign-accented speech, but how and where this is realised across studies has been inconsistent. Furthermore, no studies have set out to investigate component latency or onset latency differences. Component latency here refers to the duration of the ERP effect from onset to offset, while onset latency refers to the time-point (measured from the time-locking point) at which the difference between conditions begins to be significant [7]. The above studies all employed pre-defined time-windows or time-windows based on visual inspection of the waveforms when analysing ERP data, making it impossible to observe temporal differences between conditions. While a difference in onset latency was observed between N400s in [6], this was essentially a side-effect of employing consecutive N400 and P600 time windows. If an analysis method were used that enabled statistical comparison of component latency and onset latency more deliberately, one might expect to observe more fine-grained differences between native and foreign-accented speech on these temporal dimensions. More specifically, given that lexical retrieval seems to occur more slowly in foreign-accented than native speech [13], one might expect N400s for foreign-accented speech to exhibit a longer onset latency than their native-speech counterparts, as was found on a gross scale in [6].

Another limitation shared by these previous studies is that all were conducted in cities or states with relatively small immigrant populations. Grey and van Hell [6] was conducted in central Pennsylvania, USA, a state in which 94% of the population are American-born [14]; Hanuliková et al. [1] was conducted in Nijmegen, The Netherlands, where 75% of the population are at least second-generation Dutch [15]; and Romero-Rivas et al.[5] was conducted in Barcelona, where 86% of the population are born in Spain [16]. Note that one could argue that, because Catalan and Spanish are both official languages of Barcelona, it may be assumed that residents of Barcelona are regularly exposed to Catalan-accented Spanish. However, we do not consider Catalan-accented speakers of Spanish to be ‘foreign-accented’ for the purposes of our study, due to Catalan’s official-language status (i.e., it is not ‘foreign’). Furthermore, in Romero-Rivas et al.’s [5] study, Catalan-accented Spanish was not included as one of the foreign accents under consideration. It is therefore possible that, in the previous studies, unfamiliarity with foreign-accented speech may have affected participants’ neural responses, and that listeners with more extensive experience with foreign-accented speech may show a different pattern of results.

The first aim of the current study was therefore to compare the ERP effects elicited by semantic violations in native and foreign-accented speech (a) using a method that would allow us to deliberately examine differences in component latency and onset latency, as well as amplitude and topography, (b) in a context where participants were likely to be more familiar with foreign-accented speech. Sydney, Australia, was an ideal location for this, as only 58% of Sydney’s population is born in Australia, and 40% of the population is bilingual [4], suggesting that foreign-accented speech is widespread. We hypothesised that, as in previous studies, semantic violations produced by both speakers (native and foreign-accented) would trigger an N400 effect, but that these may show different characteristics. In particular, we expected a possible difference in the N400 effect amplitude between native and foreign-accented speech, and a potentially broader distribution and longer onset latency for N400s in response to foreign-accented speech.

Individual variability in the N400

Most previous studies investigating the neural correlates of semantic processing in foreign-accented speech have not considered potential sources of individual variability among participants. Yet, the N400 is subject to extensive individual variability [17], which can be attributed to specific characteristics of the listener. For example, the characteristics of N400s observed in response to semantic violations have been shown to covary with participant age [17] and working memory capacity [18,19]. Also, the amplitude of N400s observed in response to speaker incongruities (utterances in which the semantic content of the sentence is plausible in itself but rendered implausible by the identity of the speaker, such as the sentence ‘I think I might be pregnant’ uttered by a male voice) has been shown to vary as a function of the listener’s empathy level [20].

However, much individual variability in the N400 remains unaccounted for [17], and no study has yet described any sources of individual variability in the neural correlates of processing foreign-accented speech. This is despite behavioural evidence that skill in processing foreign-accented speech is indeed highly variable, and that this variability can, to some extent, be accounted for by participant characteristics [21,22]. While Grey and van Hell [6] attempted to account for individual variability in the N400 effect elicited by foreign-accented speech by measuring language and accent attitudes, cognitive inhibition, working memory, verbal fluency, English proficiency and IQ, no correlations between these factors and N400 amplitude were found to be significant. Therefore, the second aim of the current study, was to investigate other potential sources of individual variability in the amplitude of neural responses to semantic violations in both native and foreign-accented speech. The same factors were considered across both speaker conditions to allow for comparison. These included listeners’ age, gender, working memory capacity, empathy level, how accented they considered the speech to be, and their familiarity with the foreign accent in question. By examining these individual variability factors, the current study investigated not only how speaker-related characteristics, such as a foreign accent, immediately affect semantic processing, but also how these may immediately interact with listener-related characteristics. The potential relevance of both listener- and speaker-related characteristics has implications for neurocognitive models of sentence processing [23–27]. Integration of these characteristics will highlight the immediate use of non-linguistic information during linguistic processing, which is strongly in favour of more interactive models [24–27].

We predicted that working memory capacity would influence N400 amplitude in both native and foreign-accented speech, in line with previous findings for native speech [18,19]. While no significant relationship was found between working memory capacity and N400 amplitude for foreign-accented speech in [6], we still considered it possible that such an effect might be present in our study due to existing behavioural evidence that increased working memory capacity is associated with increased skill in processing foreign-accented speech [22]. In fact, given that processing foreign-accented speech is associated with increased working memory load [28], we expected that the influence of working-memory capacity might in fact be greater when processing the foreign-accented speech than when processing native speech. This relationship between N400 amplitude and working memory capacity is thought to occur due to the ability of listeners to take sentential context into account [19]. As semantic processing necessarily requires integration into sentential context, listeners’ ability to access sentential context may affect the degree of semantic integration difficulty and thus modulate N400 amplitude.

We also predicted that participant empathy scores might influence N400 amplitude, as more empathic listeners may be more lenient towards semantic violations made by non-native speakers, thus displaying a reduced N400 amplitude (cf. [1,29] regarding morphosyntactic violations). Regarding perceived accentedness, stronger accentedness tends to be associated with lower perceived language competence [30]. For this reason, we hypothesised that participants who perceived the foreign-accented speech to be more strongly accented might be more lenient towards the foreign-accented speaker’s errors, and thus show smaller N400 amplitudes in response to the foreign-accented speaker. Finally, participants’ familiarity with the foreign accent was expected to affect the amplitude of the N400 observed for foreign-accented speech, but not native speech. We hypothesised that participants who were more familiar with the accent (determined as those who were able to correctly identify the accent of the speaker, cf. [6]) might also expect semantic errors from the foreign-accented speaker and thus also show a smaller N400 amplitude (cf. [1,29] regarding morphosyntactic violations and [30] regarding phonological mapping).

Participants

Thirty monolingual native speakers of Australian English (13M, 17F) aged 18 to 33 years (Mean = 22 years, SD = 5.3 years) participated in the study. All lived in the greater Sydney area and had parents who were also native speakers of Australian English. The language background of the participants’ parents was controlled in order to avoid effects of exposure to foreign-accented speech (including other dialects of English) during language acquisition, which has been shown to influence speech perception later in life [31]. Participants reported no hearing, language, neurological or cognitive impairments. All were right-handed according to the Edinburgh Handedness Inventory [32]. Data from a further nine participants were excluded from analysis due to technical error (n = 4), having at least one parent who was not a native speaker of Australian English (n = 3), or having a diagnosis of depression or autism spectrum disorder (n = 2). All participants received either course credit or financial reimbursement. The Macquarie University Human Research Ethics Committee approved the experimental methods used and written informed consent was obtained from all participants before testing.

Materials

Preparation of stimuli.

Twenty-five pairs of sentences of the form “The [noun] [verb] the [noun]” were created, each pair containing a semantically congruent sentence and a corresponding semantically incongruent sentence (these pairs were termed the ‘semantic sentences’; see S1 Table for the full set of sentences). All verbs were CVC monosyllables and were in the present tense, while all nouns were disyllabic and had initial stress to create a consistent prosodic pattern across sentences. The sentences did not contain any consonant clusters except those created by the addition of the third person singular -s where necessary. This was to ensure that the target words were maximally intelligible when produced by the foreign-accented speaker (a native speaker of Mandarin): Consonant clusters, particularly in word-final (coda) position, are difficult to produce for speakers of Mandarin-accented English [33], however it was ensured that any third person singular -s occurring in a cluster was indeed produced in the stimuli. All content words in the sentences had a log frequency greater than 3.0 in the SUBTLEX-UK database [34]. All subject nouns were animate (the canonical animacy class for subjects) and consistent thematic roles were maintained across sentences (all subjects were agents, all objects were patients).

Each semantically congruent sentence was recorded by two female speakers. The native speaker was a 24-year-old monolingual speaker of Australian English. The foreign-accented speaker was a 20-year-old native speaker of Mandarin from Beijing who had been living in Australia for two years. This speaker was selected from among three candidates as being the one with the most canonical yet intelligible Mandarin accent. Sentences were elicited from the native speaker using a question and answer method to maintain consistent prosody and to ensure that the final word of each sentence was focused. As all semantic violations occurred sentence-finally, this use of focus served to maximise the salience of the violations. The question-and-answer elicitation method was too challenging for the foreign-accented speaker, so the foreign-accented sentences were elicited by having her imitate a native Australian English speaker (the first author). The two speakers differed in speech rate: Sentences were significantly shorter in duration when produced by the native speaker (Mean = 1.89 s, SD = 0.17 s) than the foreign-accented speaker (Mean = 2.07 s, SD = 0.17 s) according to a paired samples t-test (t(49) = 9.24, p < .001). However, this was unsurprising as native speakers of Mandarin have previously been found to have a slower speech rate when speaking English than native English speakers [35].

Sentences with semantic violations were created by splicing the semantically congruent sentences at the offset of the verb and appending the incongruent object from another semantically congruent sentence (see Figs 1 and 2). Splicing was conducted using Praat software (Version: 5.4.08) [36] and was used to prevent the appearance of phonetic cues to semantic incongruence occurring before the overt violation [37,38].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Diagram showing splicing locations for sentences containing semantic violations.

https://doi.org/10.1371/journal.pone.0207452.g001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Waveform and spectrogram of sample stimulus sentence “The lady bites the suitcase”.

The splicing point is shown as a dotted blue line and the time-locking point of the ERP is shown as a solid red line.

https://doi.org/10.1371/journal.pone.0207452.g002

In addition to the 100 stimulus sentences (i.e. 25 sentence pairs, with each sentence produced by both speakers), a further 200 sentences used for a different study were included in the EEG task (termed the ‘morphosyntactic sentences’; reported in [29]). These consisted of 100 sentences containing subject-verb agreement violations and 100 corresponding grammatical sentences, and contained the same phonological constraints as the sentences used in the present study (i.e. no consonant clusters except those involving the third person singular -s). One hundred filler sentences with no syntactic or phonological constraints were also included. By minimising the phonological constraints placed on the fillers, it was intended that the foreign-accentedness of the Mandarin-accented speaker would be highlighted, thus reinforcing the difference between the two speakers. All of these filler sentences were six or seven syllables in length, and thus similar in length and complexity to the target sentences. Four hundred sentences in total were therefore included in the EEG task (100 semantic sentences, 200 morphosyntactic sentences, and 100 fillers), half produced by each speaker. See Table 1 for a summary of the experimental design.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Experimental design with example sentences.

Target word is in bold.

https://doi.org/10.1371/journal.pone.0207452.t001

Procedure

During the test session, participants first completed an auditory digit span task (both forward and backward spans) to assess their working memory capacity [39], followed by a pre-task questionnaire. The questionnaire was designed to confirm that the participant fulfilled the criteria to participate in the study and to collect information to be used in the analysis of individual variability: the participant’s age and gender, and a measure of cognitive and affective empathy (the Empathy Quotient Short Form) [40].

EEG data were then collected while participants sat in a comfortable chair in front of a computer monitor in a sound-attenuated, electromagnetically-shielded room. During the task participants were presented with one of two lists of the 400 sentences via in-ear headphones. The sentences were presented in pseudorandomised order across eight blocks of 50 sentences each, with a practice block of ten additional filler sentences at the beginning of the experiment to familiarise participants with the task procedure. Each trial began with a cartoon image of an eye, presented on the screen for 2000 ms. Participants were instructed to blink whenever this image appeared on the screen. This was followed by a blank screen randomly varying in duration between 1000 and 1200 ms. A fixation cross then appeared on the screen for 500 ms before the onset of the auditory stimulus sentence and remained on the screen until 1000 ms after the completion of the sentence (see Fig 3).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Schematic diagram of the trial structure.

https://doi.org/10.1371/journal.pone.0207452.g003

Participants were presented with five comprehension questions per block at irregular intervals to ensure they were attending to the sentences (40 questions in total). Questions always referred to the sentence immediately before and responses were recorded using a button box. If participants scored below 34 out of 40 (85% accuracy), they were considered to have been insufficiently attentive during the task and were excluded. However, all participants performed at or near ceiling for comprehension accuracy, with a mean score of 39.5 out of 40 (98.7%). No participant scored below 37 out of 40 (92.5%). Therefore no participants were excluded for failing to attend to the task. Additionally, a short game of ‘Snake’ (Gremlin Industries, USA, 1976) controlled by the button box appeared five times per block for participants to play in order to maintain their attention throughout the experiment.

Data acquisition

Continuous EEG data were recorded using Curry software (Version 7; Compumedics Ltd., USA) from 64 Ag/AgCl scalp electrodes mounted on an electrode cap (Easycap, Brainworks, GmbH) according to the International 10–10 system [41] (Fpz, Fz, FCz, Cz, CPz, Pz, Oz, Fp1/2, AF7/3/1/2/4/8, F7/5/3/1/2/4/6/8, FT7/8, FC5/3/1/2/4/6, T7/8, C5/3/1/2/4/6, M1/2, TP7/8, CP5/3/1/2/4/6, P7/5/3/1/2/4/6/8, PO7/5/3/4/6/8, O1/2). Ground was located at AFz. Electrooculographic activity was measured from electrodes placed above and below the right orbit and at the outer canthus of each eye. Electrode impedances were kept below 10kΩ. Electrical activity was recorded from both mastoids, with the left mastoid as the on-line reference. The EEG signal was digitised at a sampling rate of 1000 Hz and filtered with a .05–100 Hz bandpass filter using a Neuroscan SynAmps2 DC Amplifier (Compumedics Ltd., USA).

The EEG recording lasted between one and one and a half hours for each participant, including breaks. After the EEG recording, participants completed a post-task questionnaire in which they were asked to rate the strength of the foreign-accented speaker’s accent on a seven-point scale from ‘No foreign accent’ (1) to ‘Very strong foreign accent’ (7), and to identify the nationality of the foreign-accented speaker. In all, the testing session for each participant lasted approximately two and a half hours.

Data processing

EEG data were processed in MATLAB (Version R2015a, The MathWorks Inc., Natick, MA, US) using the Fieldtrip toolbox (Version 20160719) [42]. Epochs of approximately 5500 ms were extracted from the continuous data stream (from 1000 ms before the presentation of the fixation cross to 1500 ms after the offset of the audio file for each trial). An independent component analysis was then performed on the data to remove artefacts caused by eye-blinks, saccades, and pulses [43]. The data were re-referenced off-line to the average of the mastoids. A baseline correction was then performed, with the 200 ms before the time-locking point of the ERP as the baseline period. (Before baseline correction, a cluster-based permutation test [44] was performed on the data, which showed no significant differences between conditions during the baseline period, confirming its suitability). A zero-phase fourth-order Butterworth IIR low-pass filter of 35 Hz was applied to the dataset, with a 12 dB per octave slope. Epochs were then redefined as the period from the beginning of the baseline period to 1500 ms after the time-locking point (the onset of the object noun).

Trials with remaining artefacts, defined as those epochs in which the amplitude variance was greater than 1000 μV and/or the maximum z value was greater than 4, were then removed from the data set. In total, 227 of the 3000 trials for this study were rejected (7.6%). These were evenly distributed across sentence types. Once trials containing artefacts had been rejected, the data were averaged within sentence types for each participant, then a grand average for each sentence type was calculated across participants.

Statistical analysis

Cluster-based permutation tests (CBPTs) [44] were used to compare neural responses between conditions. To conduct a CBPT, a t-test is first performed on the amplitude values of the two waveforms to be compared at each individual sampling point. Sampling points at which the difference between the waveforms is found to be significant (p < 0.05, two-tailed) are then formed into clusters based on both temporal and spatial adjacency, and the polarity of the effect (i.e. t-values associated with sampling points in a single cluster must be either all positive or all negative). Each electrode has a mean of 6.27 ‘neighbour’ electrodes (range = 2–8) which are considered spatially adjacent. Cluster-level statistics are then calculated by adding together all t-values within the cluster. The largest of these cluster-level statistics are then tested for significance against clusters of samples selected using 1000 random partitions. CBPTs are therefore able to provide information about the polarity, amplitude, latency (duration from effect onset to offset), onset latency (time point of the effect onset), and topographical distribution of any ERP effect. They are also recommended by Luck and Gaspelin [45] as a method which minimises the possibility of Type I error in ERP analysis, compared to the traditional approach of using pre-defined time-windows and multi-level ANOVAs for analysis.

Three CBPTs were performed on the data across the 1000 ms following the time-locking point of the ERP. The first compared neural responses to semantically congruent and incongruent sentences produced by the native speaker to test if an N400 effect was elicited by the native speaker’s semantic violations. The second CBPT compared semantically congruent and incongruent sentences produced by the foreign-accented speaker to test if an N400 effect was elicited by this speaker’s violations. Finally, neural responses to semantically incongruent sentences for each speaker were subtracted from responses to that speaker’s semantically congruent sentences. These difference waves were then compared using a third CBPT to examine whether the characteristics of the N400 effects differed between the two speaker conditions, i.e. an interaction between Speaker and Congruity.

Responses to semantic violations in native vs. foreign-accented speech

The first aim of the current study was to test whether semantic violations in foreign-accented speech elicited different ERP effects from those in native speech in a linguistic environment where foreign-accented speech was relatively common, using a method that could account for all characteristics of an ERP component. We hypothesised that the same ERP effect, an N400 effect, would be found for both speakers’ semantic violations, but that these N400 effects would differ in terms of their amplitude, topography, and/or onset latency. Semantically-incongruent sentences did elicit a significantly larger negativity than semantically-congruent sentences in both speaker conditions. The larger negativity in the native-speaker condition showed a broad distribution with a centro-parietal maximum while the larger negativity in the foreign-accented-speaker condition showed a centro-parietal distribution (see Fig 5). These differences are interpreted as N400 effects, motivated by their polarity and scalp distribution. This supports our hypothesis that an N400 effect would be present for semantic violations produced by both speakers.

The CBPTs performed separately for the two speakers appeared to show differences in topography, component latency and onset latency between the two speaker conditions: The N400 for foreign-accented speech seemed less broadly distributed (see Fig 5 above), shorter in latency (93 vs 567 ms), and later in onset latency (564 vs 364 ms) than that for native speech. To test whether these differences were significant, a CBPT was conducted on the difference waves of the two speaker conditions, but it revealed no significant differences. That is, any apparent difference between the two N400s did not reach statistical significance. This seeming inconsistency may be due to the sensitivity of CBPTs to signal-to-noise ratio. Data in the foreign-accented-speaker condition could potentially be noisier than those of the native-speaker condition (for example, if inter-listener variability in responses was greater for foreign-accented than native speech), resulting in an apparently smaller effect for the foreign-accented speaker when individual CBPTs were performed. However, without a suitable metric to quantify data noisiness across conditions, this is a speculative explanation.

These results contradict our hypothesis that the N400s observed for violations produced by the two speakers would differ in their amplitude, topography and/or onset latency, as no differences on any dimension were found. This also goes against the results of other studies comparing neural responses to semantic violations in native and foreign-accented speech, which all found differences between the two speaker conditions [1,5,6]. There are two potential explanations for the lack of a difference in the current study: the statistical methods used, and the linguistic environment in which the study was conducted.

The first potential explanation for the lack of difference between the N400s for the two speakers relates to the use of CBPTs for statistical analysis. The CBPT is a conservative statistical method which greatly minimises the risk of finding false positive results in comparison to the often-employed method of analysing ERP data with multiple ANOVAs across pre-defined time windows [45]. As pre-defined time windows or time-windows based on visual inspection of the data and multiple ANOVAs and post-hoc tests were used in each of the studies finding significant differences between N400s for native and foreign-accented speech [1,5,6], it is possible that these less-conservative statistical methods were responsible for the effects found in these studies, as the possibility of Type I error was greatly inflated. The inconsistent nature of the differences found between N400 effects in the previous studies lends some support to the proposal that the effects may be spurious. Furthermore, according to simulations carried out by Luck and Gaspelin [45], the probability of at least one Type I error occurring in an ERP analysis based on factorial ANOVAs may be as high as 96%, depending on study design. Alternatively, it is possible that the linguistic environment that participants in the current study were exposed to may have been responsible for the lack of effect found. Evidence for this comes from the analysis of individual variability, discussed below.

Individual variability in N400 amplitude

The analysis of individual variability in listeners’ neural responses revealed several factors which significantly influenced N400 effect amplitude. Note that, as we were interested in factors that affect effect amplitude, rather than component amplitude, we will focus here only on factors which interacted with Congruity. Three factors were found to influence N400 component amplitude depending on the congruity of the sentence: familiarity, working memory and age.

Implications

As mentioned previously, foreign-accented speech is commonly encountered in everyday life, and thus must be accounted for by models of sentence processing. However, the current study has demonstrated that the impact of a foreign accent on semantic processing is not an all-or-nothing phenomenon: Gradient effects of foreign-accented speech may be observed, mediated by the characteristics of individual listeners such as familiarity, working memory capacity and, potentially, age. This suggests that models of semantic processing, and sentence processing more broadly, must not only account for processing changes due to characteristics of the speaker (e.g., foreign-accentedness), but also individual processing differences due to characteristics of the listener, and interactions between these factors. From this perspective, the results of the current study provide support for more interactive models of sentence processing [24–27]. These models can account for the integration of multiple sources of not only linguistic information during sentence processing, but also non-linguistic information, including socio-indexical information about the characteristics of the speaker, as is the case here.

By conducting the current study in a linguistic environment where foreign-accentedness is widespread, we also observed a different pattern of results to previous studies conducted in locations were foreign-accentedness was less common [1,5,6]. This provides some evidence that the ambient linguistic environment can have a considerable impact on the results obtained in a sentence processing study using ERPs. This suggests that the linguistic environment must be taken into account in future studies of sentence processing.

A direction for future research also arises from a potential limitation of the current study, where only a single native speaker and a single foreign-accented speaker provided the stimuli. While this approach was chosen to increase the consistency of the stimuli, minimise unnecessary variability in neural responses due to speaker-specific effects, and align with the procedure used in previous studies [1,6], the result is that we cannot definitively state that the difference in responses to the two speakers found among the less-familiar listeners was due to the foreign-accentedness of the speaker per se. It is possible that some idiosyncratic property of the speech of one or both speakers is in fact responsible for the effect. We believe this is unlikely, as the two speakers were chosen for their canonical native Australian and Mandarin-accented speech respectively, however further research involving stimuli produced by a greater number of native and foreign-accented speakers would be valuable to provide supporting evidence for our findings.