Positionality statement

The authors of this paper are members of the Speech Perception and Production Laboratory at the University of Oregon, including faculty, postdoctoral fellows, graduate students, and undergraduate students. The authors come from diverse linguistic, socioeconomic, and racial backgrounds and have had diverse experiences, which impact their perspectives, biases, and research interests. We refer the reader to this webpage https://www.speechperceptionproductionlab.com/positionalitystatments, where each of our positionality statements is available.

Problems with “non-native”

“Non-native,” a term used frequently in linguistics and psychological research, more generally refers to something that is not endemic to a region or place. This is ironic within a linguistic context as most of the dominant languages in the world have their dominant status due to dissemination and colonization of places and peoples (Cooper, Reference Cooper1982). The term has been used in linguistics for decades; Leonard Bloomfield used it to describe the first language that someone acquires (Reference Bloomfield1933). The term remained largely uninterrogated and broadly accepted as a useful delineator between speakers. The definition resulted in the dichotomized nature linguists use to describe the differences between variable ways of speaking (Chomsky, Reference Chomsky1965), and while this view evolved to include not only grammatical competence but also communicative competence (Chomsky, Reference Chomsky1980), the prevailing notion became that a native speaker was the only reliable source for linguistic data. The notion of “non-nativeness” is exacerbated by nomenclature, but, importantly, the issues of discrimination and mistreatment of “non-native” speakers persist no matter what label is used, particularly within the frame of experimental design.

Experimental research in psycholinguistics is a lucrative area for the language learning industry, which has become a space that further perpetuates the native/non-native binary. The funding that goes into these sorts of programs could be guiding the kinds of questions researchers are asked to investigate (Kilman et al., Reference Kilman, Zekveld, Hällgren and Rönnberg2014; Tamminen et al., Reference Tamminen, Peltola, Kujala and Näätänen2015). The mass interest in language learning covertly reinforces the need to teach speakers rather than gaining a better understanding of how listeners perceive speech (Flores & Rosa, Reference Flores and Rosa2015; Ramjattan, Reference Ramjattan2019). Thus, the communicative burden is typically placed on those considered non-native, insinuating the need to be very proficient in the language of wider communication to be successful. Baese-Berk et al.’s (Reference Baese-Berk, McLaughlin and McGowan2020) review on perception of non-native speech discusses the many factors that impact its perception: The relationship one has between their native language and people who speak it, their relation with people who speak a non-native variety, as well as perception of non-native speech drives the kinds of research questions people ask.

To reframe non-native speech as unfamiliar speech would benefit the biggest goal of communication: to be mutually understood in terms of intelligibility, no matter what one’s L1 is. However, this reframing also has more specific benefits. It could help researchers clarify precisely what questions they are addressing and incorporate these questions in a broader literature. It is clear from a wide body of psycholinguistic literature that listeners process familiar talkers (i.e., those talkers they have heard before either through life experience or through experimental exposure) differently from unfamiliar talkers (i.e., talkers they have not heard previously; e.g., Levi et al., Reference Levi, Winters and Pisoni2011; Nygaard et al., Reference Nygaard, Sommers and Pisoni1994; Nygaard & Pisoni, Reference Nygaard and Pisoni1998), and this benefit has been shown to occur for both talkers with accents that the listener has prior exposure to (e.g., Nygaard & Pisoni, Reference Nygaard and Pisoni1998) and accents that a listener has little to no experience with (e.g., Levi et al., Reference Levi, Harel and Schwartz2019). Given this understanding that unfamiliar speech of various types is challenging for listeners, we are able to acknowledge communicative difficulties that affect all people, regardless of language background. We propose harnessing the power of the dichotomies that currently exist (i.e., native vs non-native) to shift how we as experimentalists are framing non-native speakers.

Framing non-native speech as unfamiliar speech shifts the communicative burden from the speaker to the listener, who often occupies a position of power. This framing, however, maintains a binary, rather than recognizing that there is likely a continuum of familiarity. For example, one may be much more familiar with the voices and speech of their close family and friends than with acquaintances; however, this same individual is likely much more familiar with the voices of these acquaintances than with a speaker from a language background they have not previously encountered at all. Therefore, we would expect familiarity to be a gradient construct, likely co-varying with experience, that may have gradient results on psycholinguistic work. Further, this familiarity is likely to vary as a function of context. For example, an individual living in the American Southwest may not speak Spanish themselves but is more likely to have encountered Spanish, varieties of English influenced by Spanish, or terms that have come out of Spanish/English contact than varieties of English influenced by, say, Khmer ^{Footnote 1} . There is a growing body of psycholinguistic work that considers factors that impact listeners’ perception of and adaptation to unfamiliar speech; this work does shift the communication burden in a practical way (e.g., Kutlu et al., Reference Kutlu, Tiv, Wulff and Titone2021).

Another key issue is what precisely is unfamiliar to a listener in these scenarios. As an anonymous reviewer noted, a listener is likely to be unfamiliar with a wide array of cues ranging from phonetic to lexical to cultural. Which of these cues are most relevant for a listener is likely to depend on the precise research question being investigated. As we think about research design going forward, definitions of familiarity should be explicitly stated according to the questions (and answers) each study is hoping to address. It is also important to consider the power of social information when constituting unfamiliarity, as social information can dictate to a listener that they are hearing something in the speech signal that is not acoustically there, from something as simple as a fabrication regarding where the speaker is from or a picture of a face (Niedzielski, Reference Niedzielski1999; Rubin, Reference Rubin1992).

Using familiar and unfamiliar as framing devices provides an opportunity to consider perception of non-native speech not as a specific problem to be solved that is distinct from other issues in speech perception but rather as a specific instance of a broader problem which is addressed throughout related literature—how a listeners’ familiarity with a particular voice, accent, or dialect impacts their perception. Instead of focusing exclusively on properties of the talker that make speech perception challenging, we can instead point to the fact that listeners face challenges across a range of listening conditions, and this particular condition may just be one instance of this challenge. This framing shifts the communicative burden from speaker to listener, benefitting all speakers. Still, reframing does not solve the basic problem we see in this work. The issue at hand is not fundamentally about terminology but more centrally about the biases brought to bear in the research setting, and how we can use the existing dichotomies to our advantage in psycholinguistic research specifically. We know that the discerning nature of humans can lead to discrimination and bias (Weissler, Reference Weissler2022), and therefore, there needs to be dedicated consideration by the psycholinguistics community, and by people more broadly, about the powerful influence of ideology in every facet of our lives, including in how we design our scientific experiments.

Factors affecting the perception of non-native talkers

We see in many studies that listening to speech with unfamiliar accents requires more effort than speech with familiar accents, even if the speakers are fully intelligible (e.g., McLaughlin & Van Engen, Reference McLaughlin and Van Engen2020). One of the claims we see in the literature addressing perception of native and non-native talkers is that speech from non-native talkers is inherently more difficult to perceive (Bent & Frush Holt, Reference Bent and Frush Holt2013; Leikin et al., Reference Leikin, Ibrahim, Eviatar and Sapir2009; Munro & Derwing, Reference Munro and Derwing1995). However, knowing what makes speech more difficult to perceive is complex because not all difficult listening conditions are created equal (see Mattys et al., Reference Mattys, Davis, Bradlow and Scott2012 for a review). For example, speech presented in noise is processed differently from accented speech (Adank et al., Reference Adank, Nuttall, Banks and Kennedy-Higgins2015), and even different types of noise (such as white noise vs. cocktail party babble) seem to affect speech recognition in different ways, at different times in processing (Mattys et al., Reference Mattys, Davis, Bradlow and Scott2012; McLaughlin et al., Reference McLaughlin, Baese-Berk, Bent, Borrie and Van Engen2018). Furthermore, other types of unfamiliar speech, like regional accents and dialects, or speech from people with dysarthria and other speech disorders are also difficult to perceive but in ways that are distinct from non-native speech (Bent et al., Reference Bent, Baese-Berk, Borrie and McKee2016). Researchers have also pointed out the wide range of variation in individuals’ ability to understand speech in noise, accented speech, or disordered speech (Bent et al., Reference Bent, Baese-Berk, Borrie and McKee2016).

In addition to how speech in general is perceived under different conditions, there are both cognitive and social factors that affect how we perceive unfamiliar accents. Cognitive factors include the size of our receptive vocabulary (i.e., the number of words we understand but not necessarily the number of words that we use in our own speech) and our working memory capacity (e.g., Adank et al., Reference Adank, Evans, Stuart-Smith and Scott2009; Banks et al., Reference Banks, Gowen, Munro and Adank2015, Reference Bent, Baese-Berk, Borrie and McKee2016; McLaughlin et al., Reference McLaughlin, Baese-Berk, Bent, Borrie and Van Engen2018). Social factors, on the other hand, include how we associate accents with faces, voices, and identity markers, often races: when our assumptions about what type of person has what kind of accent are violated, we find the speech more difficult to perceive (Hanulíková, Reference Hanulíková2021; Hanulíková et al., Reference Hanulíková, van Alphen, van Goch and Weber2012; Kutlu, Reference Kutlu2020; Kutlu et al., Reference Kutlu, Tiv, Wulff and Titone2022; McGowan, Reference McGowan2015; Rubin, Reference Rubin1992; Vaughn, Reference Vaughn2019). We discuss these factors in more detail below.

In speech perception research, we use the term “intelligibility” to refer to the extent to which listeners can correctly recognize the speech stimuli they hear, often as measured by the number of correctly transcribed words from an auditory stimulus. This measure represents how accurately a speech signal is perceived, which is a beneficial framework to use when considering unfamiliar accent perception. A number of studies have relied on intelligibility measures to discover the effects of adverse conditions on speech processing due to its ease of interpretation and possibility of cross-study comparison. Multiple studies have found that the size of a participant’s receptive vocabulary is a good predictor of their intelligibility score. Listeners with larger receptive vocabularies tend to provide higher intelligibility scores, while the opposite is true for participants with smaller receptive vocabularies.

In experiments where information about where a speaker is from is manipulated, listeners familiar with the region and accent perceive the same recorded vowel productions differently depending on the social or dialect group they believe the speaker to be a part of (Hay et al., Reference Hay, Warren and Drager2006; Hay & Drager, Reference Hay and Drager2010; Niedzielski, Reference Niedzielski1999). Perceived social group identity as indexed by the physical appearance of a speaker has also been shown to affect listener perception through the use of matched-guise tasks or pairing images of speakers’ faces with audio. Rubin’s (Reference Rubin1992) classic finding that the same recording would be rated as more accented when paired with images of Chinese faces than when paired with images of white faces has been investigated further in recent studies. For example, speech paired with South Asian faces in a matched-guise task was rated as more accented than speech paired with white faces in a mostly monolingual setting (Kutlu, Reference Kutlu2020) but not when participants are located in an area with more exposure to multilingual speakers (i.e., Montréal; Kutlu et al., Reference Kutlu, Tiv, Wulff and Titone2022). Babel and Russell (Reference Babel and Russell2015) similarly found that speech paired with photos of Chinese faces led to lower intelligibility scores and higher accentedness ratings than speech paired with white faces. They found these results despite another group of listeners being unable to reliably distinguish between the Chinese and white talkers when the speech was presented without photos, even though listeners resided in a Canadian neighborhood with a large multilingual and multicultural population.

Similar evidence of a complex interaction between listener experience with and expectations regarding speech varieties, speakers, and contexts and the way speech is evaluated has also been found in matched-guise tasks. Speech is transcribed less accurately when an image of the purported talker did not match the listener’s expectations (e.g., a white face paired with Mandarin-accented utterances; McGowan, Reference McGowan2015). Listeners have also been shown to transcribe speech more accurately when they are given information about the speaker’s accent than when they receive no information about the accent (Vaughn, Reference Vaughn2019).

Neuroimaging studies have also shown that listeners’ processing is modulated by how the speaker is perceived. When listening to speech, electrophysiological responses to morphosyntactic errors (Hanulíková et al., Reference Hanulíková, van Alphen, van Goch and Weber2012) and semantic anomalies (Hanulíková et al., Reference Hanulíková, van Alphen, van Goch and Weber2012; Romero-Rivas et al., Reference Romero-Rivas, Martin and Costa2015) are elicited when the speech comes from a native speaker but not when it comes from a non-native speaker. Listeners also show differences in the synchronization of their neural oscillations when they listen to speech in their native language, a foreign language that they have learned, and a language completely unknown to them (Jin et al., Reference Jin, Díaz, Colomer and Sebastián-Gallés2014; Pérez et al., Reference Pérez, Carreiras, Gillon Dowens and Duñabeitia2015). Studies have also found a neural response bias in favor of the listener’s own accent, with decreased neural response to accents of social groups the listener is not a part of (Bestelmeyer et al., Reference Bestelmeyer, Belin and Ladd2015).

The perceptual effects that we see in behavioral and neuroimaging studies occur with unfamiliar accents and listeners’ perception is shaped by co-occurring social information, like photos and descriptions of racioethnic traits. This points to the factors affecting perception arising from both properties of the listener and the speech signal itself. In addition, facets of listener experience that may also impact perception of unfamiliar accent include factors as disparate as musical experience (Kraus & Chandrasekaran, Reference Kraus and Chandrasekaran2010; Kraus et al., Reference Kraus, Slater, Thompson, Hornickel, Strait, Nicol and White-Schwoch2014; Moreno et al., Reference Moreno, Marques, Santos, Santos, Castro and Besson2009; Musacchia et al., Reference Musacchia, Sams, Skoe and Kraus2007; Qin et al., Reference Qin, Zhang and Wang2021; Zhao & Kuhl, Reference Zhao and Kuhl2016; Zhao et al., Reference Zhao, Llanos, Chandrasekaran and Kuhl2022) and motivation (Gardner & Lambert, Reference Gardner and Lambert1959; Gardner et al., Reference Gardner, Tremblay and Masgoret1997; Saito et al., Reference Saito, Dewaele, Abe and In’nami2018; Tsang, Reference Tsang2022). The impacts of such diverse factors complicate approaches to measuring the perception of unfamiliar speech; researchers must take into account the contribution of the listener to speech perception tasks and assess whether the questions we ask and empirical methods we employ accurately address all of the factors at play in perception of different speech varieties. With this point in mind, we must ask whether the tasks we use to measure perception of non-native speech are, in fact, accurately measuring what we believe they are. We discuss this point in the following section.

What are our psycholinguistic measurements capturing?

The selection of which task to use in a study is a complex decision, as the type of task a participant is asked to do affects how they interact with both the task and the materials used in the experiment. For example, listeners demonstrate different performance if they are asked to do a simple discrimination task compared to a more complex task requiring listeners to compare stimuli across trials (e.g., Pisoni & Lazarus, Reference Pisoni and Lazarus1974). The assumptions about a task also affect how results are interpreted. All of these factors together mean that we must know whether our tasks are measuring what we think they are. We are choosing to focus on transcription tasks, but similar questions can (and should) be asked of other tasks.

Transcription tasks provide an intelligibility score based on the number of words a listener correctly transcribes from each utterance in a stimuli set. However, apart from demonstrating the accuracy of response, the metric does not provide an insight into underlying causes of listeners’ performance. Moreover, even when participants are able to accurately transcribe all of a talker’s speech, there is evidence that they exert additional effort to understand non-native talkers than with native talkers (Brown et al., Reference Brown, McLaughlin, Strand and Van Engen2020; McLaughlin & Van Engen, Reference McLaughlin and Van Engen2020). This discrepancy between the required listening effort and intelligibility suggests that there are additional factors that an intelligibility score is not capturing. “Listening effort” itself is another complex construct made up of multiple factors, which are difficult to tease apart (McGarrigle et al., Reference McGarrigle, Munro, Dawes, Stewart, Moore, Barry and Amitay2014; Van Engen & Peelle, Reference Van Engen and Peelle2014). The exact nature of the reasons behind increased effort when listening to non-native speech remains unclear. It is likely an indication of deviations between L1 and L2 language patterns that require more cognitive resources to correctly map those representations (Brown et al., Reference Brown, McLaughlin, Strand and Van Engen2020). Regardless of the source of difficulty, it is important to highlight the need to better understand how non-native speech affects perception and the inability of the intelligibility measure in isolation to provide the full picture of all listening challenges (Baese-Berk et al., Reference Baese-Berk, McLaughlin and McGowan2020, under review).

Because intelligibility tasks are not very sensitive, alternative or additional measures should be considered as important means to capture the nuance of what exactly is “difficult” about perceiving accented speech. Both additional research methods and additional measures available to researchers using intelligibility tasks can provide insights on the complex nature of this area of perception. For example, instead of using the percentage of correctly transcribed words alone to investigate perception of various accents, researchers can classify the types of errors that listeners made during transcription (Winn & Teece, Reference Winn and Teece2021). As Winn and Teece point out, participants may have similar intelligibility scores despite committing different errors—and different types of errors that suggest different amounts of cognitive effort are scored the same way in a standard intelligibility measure. Errorful responses that are more plausible than the intended sentence may, in fact, require less cognitive effort than providing the correct response. Even when intelligibility is fully achieved, correct responses may demand various degrees of cognitive effort. Pupillometry data reflect processing of non-native speech in a more nuanced way by capturing an increase in listening effort when intelligibility is at ceiling, which would otherwise remain undetected in standard transcription tasks (McLaughlin & Van Engen, Reference McLaughlin and Van Engen2020).

In addition to qualitative examinations of error type, it may be useful to incorporate passive physiological measures that are associated with processing difficulty, such as pupillometry, skin conductance, and various types of neural responses. Due to the level of sensitivity to online processing they afford, these passive measures collect additional information about listener effort when processing speech. Gathering fine-grained information about listener processing is vital. Studies using these passive measures have found that variation in listener-related factors such as level of experience with the accent being perceived (Brown et al., Reference Brown, McLaughlin, Strand and Van Engen2020; Porretta & Tucker, Reference Porretta and Tucker2019) and listener personality traits (Francis et al., Reference Francis, Bent, Schumaker, Love and Silbert2021) affect processing effort during perception of non-native speech.

Passive measures have the additional advantage of demonstrating the nature of online processing, before the transcription is provided and analyzed in addition to the accuracy of the transcription or other offline measures. Utilizing multiple offline measures may be just as informative as online measures. For example, researchers also use subjective (e.g., self-report questionnaires; Koelewijn et al., Reference Koelewijn, Zekveld, Festen, Rönnberg and Kramer2012) and behavioral measures (e.g., response times; Munro & Derwing, Reference Munro and Derwing1995; Floccia et al., Reference Floccia, Butler, Goslin and Ellis2009) to provide additional insight into listening effort exerted during the perception of non-native speech. These approaches used in combination with transcription tasks can make our analysis more informative: while relying on binary responses (“correct” or “incorrect”) limits us to interpreting only the result of the perception, adding other measures helps us better understand the cognitive processes determining certain behavioral response and get a better sense of the factors that combine to form the construct of intelligibility.

By using measures that describe what drives listeners’ behavior and reflect underlying cognition during speech perception, we focus on the listener and the effort they do or do not exert. Looking at intelligibility from this perspective has the potential to bring us closer to understanding how to minimize the cognitive load of the listener and improve their perceptive abilities without placing the communicative burden entirely on the speaker and the characteristics of their speech. Furthermore, while the listening subject is often perceived as white (Flores & Rosa, Reference Flores and Rosa2015), we believe the issues of communicative burden also harm white speakers, particularly through the intersection of socioeconomic background and disability. We thus do not want to imply that “nativeness” is inherently white, rather that those who are considered native are typically those who are in power in a given society.

Recommendations for future work

Thus far, we have described the notion of non-native from various vantage points, particularly relating this term to both linguistic and experimenter biases. Here, we provide recommendations for future work surrounding how we as scientists can reframe the notion of “non-native” to help shift the communicative burden from speaker to listeners of all language backgrounds.

The issue of non-native within psycholinguistics has begun to be approached by finding more accurate ways to characterize the people whose language we are studying and working with (Cheng et al., Reference Cheng, Burgess, Vernooij, Solís-Barroso, McDermott and Namboodiripad2021), such as removing the dichotomy of native and non-native to more accurately represent speakers’ history, identity, and other continuous measures of speaker performance. This is a useful step forward, though as we described, the issue of terminology is reflective of a bigger issue: the focus of much work in this area focuses on making the non-native speaker “better,” in terms of their speech being more intelligible, more aligned with a prestigious variety, or easier to understand. However, it is crucial to understand that these terms were not created by minoritized speakers themselves and are not typically focused on minoritized speaker equality or liberation. In reality, most labels will inherently be exclusionary in one way or another, and even the most inclusive terminology does not solve the overarching cause of the problem—“native” listeners often have challenges in understanding unfamiliar accents that sometimes have little to do with the accents themselves and more to do with cognitive and social factors of the listener themselves. We suggest that using the terms that already exist can be helpful to frame experiments as how listeners in general interact with unfamiliar speech on a continuum, rather than as a dichotomy.

We also need to be cognizant of the globalized society we live in and the ideological backgrounds and socialization that our participants come from. This entails addressing real-world marginalization. One approach to address these issues is to investigate ways to encourage listeners to improve and challenge their listening abilities, which can be difficult when there is a seemingly baseline notion of validity of talkers (e.g., Chomsky’s (Reference Chomsky1965) definition of the ideal speaker being a native speaker of a language). Further, examining our understanding of the motivations for and results of learning additional languages is crucial. As discussed above, many extralinguistic factors affect how a listener perceives speech, including the listener’s own biases, expectations, and background. To more accurately represent these facets of the listener, it would be useful to collect and include more data in the analysis of the listener’s responses—for example, language background/experience questionnaires or attitudinal surveys in addition to accuracy or intelligibility scores, reaction times, or neuroimaging data.

Some steps in the direction toward equitable psycholinguistics research practice are already taking effect. There are examples of psycholinguistic work that embrace the dichotomy and shift the communicative burden to the listener. For example, Kutlu et al. (Reference Kutlu, Tiv, Wulff and Titone2021, Reference Kutlu, Tiv, Wulff and Titone2022) further the field’s understanding of perception of non-native speech, finding that listeners with less racially diverse social networks give higher accentedness judgments to non-native speakers, further suggesting the impact of social perception (or lack of exemplars) and speech perception.

Linguistic anthropology and education are related fields that also explore these problems through the lens of the white listening subject, making a concrete link between familiarity of languages and intercultural competence (Hannerz, Reference Hannerz1973; Jung, Reference Jung2010; Flores & Rosa, Reference Flores and Rosa2015). The fundamental strategy of framing linguistic knowledge as cultural competence would be beneficial for psycholinguistics to pull from when conceptualizing experiments. The framework of recognizing that all people are socialized to perceive white as right requires an interrogation of what biases we have when listening to speech. Gerald (Reference Gerald2020) suggests teaching the white perceiver to raise their communicative awareness, potentially by meeting the listeners where they are. By emphasizing that the ways we perceive can be augmented and shifted, listeners can gain a clearer sense of what they are tuning into when listening to all speech. Being clear that the perceptions we have are learned knowledge and not inherent knowledge can combat the notions that conceive of “native” speech as better than non-native. In addition to using established theories from related fields to inform our research, we can revise existing theories and frameworks within linguistics to account for the social factors that affect language processing: consider Chomsky’s shift from his 1965 work to his 1980 work considering only grammatical competence to including communicative competence as well. An ideal theory would expand further to incorporate theory from the aforementioned fields as well as speech familiarity to reflect how our linguistic knowledge changes over time.

We want to acknowledge that some tools we have mentioned in this paper might not be accessible to every researcher. As such, we suggest behavioral data collection tools that are free or inexpensive to license and flexible to use with many methodologies, such as Gorilla, PCIbex, PsychoPy/Pavlovia, and others. We would also like to suggest incorporating multiple types of data that are collected, using both passive and active measures, online and offline measures, or multiple measurements of all types. Finally, we would also like to highlight the various online recruitment platforms that exist as of this writing, which can help expand the reach of our experiments to more than the usual college student participant population and ideally make our results more broadly applicable and relevant to multiple populations.