Keeping track of language: Can monolingual and bilingual infants associate a speaker with the language they speak?

Abstract

Bilingual infants acquire languages in a variety of language environments. Some caregivers follow a one-person-one-language approach in an attempt to not ‘confuse’ their child. However, the central assumption that infants can keep track of what language a person speaks has not been tested. In two studies, we tested whether bilingual and monolingual 5-, 12- and 18-month-olds spontaneously form language-person associations. In both studies, infants were familiarized to a man and a woman, each speaking a different language, and tested on trials where they either spoke the same language or switched to a different language. In Study 1, infants only heard the speaker, and in Study 2 infants saw and heard the speaker. Bilinguals and monolinguals did not look longer to Switch compared to Same trials; there was no evidence in this task that infants form person-language associations spontaneously. Thus, our results did not support a central assumption of the one-person-one-language approach, although we cannot rule out that infants do form this association in more naturalistic contexts.

Bilingual children regularly interact with individuals speaking different languages, but relatively little is understood about how the social aspects of daily interactions might support their language learning. For monolingual learners, learning about people and about language go hand in hand. For example, infants use information about what language a person speaks to reason about people’s preferences and social affiliations (Kinzler, 2021). Similarly, bilingual adults use a person’s language preference to generate predictions for language processing (Martin et al., 2016). For bilingual infants, being clued into who speaks what language could be especially useful. One theory of how bilingual infants acquire two languages and ultimately are able to speak each language posits that they track the sounds, words and rules of each of their languages separately (Nazzi et al., 1998). In this case, being able to associate a person with the language they habitually speak could be helpful to encode information from the same language together (see also Kandhadai et al., 2017). In two studies each testing two age groups, we tested whether 5- to 18-month-olds spontaneously associate a person with the language they speak.

Bilingual Language Environments

Bilingual infants vary in terms of how they encounter their languages. One popular way to raise a bilingual infant is for two caregivers to each speak a different language to the child, also known as the “one-person-one-language” approach (Barron-Hauwaert, 2004). Early bilingual theorists hypothesized that this approach would allow children to separate their languages from the beginning, thus preventing confusion (Grammont, 1902). While more modern research has demonstrated that children can successfully acquire two languages from various family language configurations (De Houwer, 2007), it is nonetheless relevant to understand whether person-language associations are made in early bilingual language acquisition. While we are just beginning to describe bilingual infants’ learning environments (Kremin et al., 2022; Orena et al., 2020), it is probable that the people infants encounter use their languages in person-specific ways. Thus, keeping track of who speaks what languages could make it easier for infants to detect structure in their language environment.

Foundational Skills

In order to form person-language associations, infants need to be able to discriminate between people and between the languages around them. Newborns can discriminate male and female voices (Floccia et al., 2000), and at least by 1 month of age, infants can discriminate unfamiliar faces when shown static images (de Haan et al., 2001). Same-gender voice discrimination seems to develop later in the first year of life. Starting at 7 months, monolingual infants can discriminate two female voices if the voices speak the infants’ native language (Fecher & Johnson, 2018; Johnson et al., 2011). Bilingual infants may have an advantage in voice discrimination. In a set of studies, bilingual and monolingual 9-month-olds were habituated to a voice-face pairing, and at test the voice was changed. Bilingual infants were able to detect the change when they heard a familiar and an unfamiliar language, while monolingual infants only noticed the mismatch in a familiar language (Fecher & Johnson, 2019; 2022). Despite some of these nuances, overall, these studies show that infants learn to discriminate both same-gender as well as different-gender speakers early on, which is an important prerequisite for making person-language associations.

Infants can discriminate many language pairs from birth. This is well-documented in studies where infants hear long passages of speech and their reactions to a language switch are measured. Newborns’ ability to discriminate languages depends on the similarity of the pairs of languages measured (Gasparini et al., 2021). Using auditory information only, infants can discriminate rhythmically different languages like English and Tagalog at birth (Byers-Heinlein et al., 2010), and rhythmically similar languages like Spanish and Catalan at least by 4 months of age (Bosch & Sebastian Galles, 2001; Molnar et al., 2014). Infants can also rely only on visual information to discriminate languages. In these studies, 4- to 8-month-old infants see a video of a person talking without sound, and are surprised when the person starts speaking a different language (Sebastian Galles et al., 2012; Weikum et al., 2007). When given both auditory and visual information, bilingual and monolingual 4-month-old infants can discriminate rhythmically distant languages, however bilingual infants do not seem to be able to discriminate rhythmically close languages (Birulés et al., 2018). Again, while there are some nuances to language discrimination, the ability to discriminate many language pairs emerges early.

Evidence for Sophisticated Reasoning about Language in Social Contexts

While no studies have tested if infants can make spontaneous associations between a person and a language directly, there is evidence that infants are able to reason about how people they interact with use their language(s). This evidence comes from studies on mostly monolingual infants’ reasoning about social groups. These studies differ from the studies reviewed so far in that they often test infants in highly social situations and/or measure preference for groups of people. However, these studies do highlight that language use can be a salient cue for infants’ reasoning about people. For example, monolingual 10-month-old infants preferred to take a toy from a person who spoke their native language (Kinzler et al., 2007). Monolingual infants also selectively imitate speakers of their own native language (Buttelmann et al., 2013; see also Howard et al., 2014). Moreover, monolingual infants use the languages they observe others speak to reason about their social relationships (Liberman et al., 2017). Observing a person speaking a familiar or unfamiliar language also affects whether infants expect object labels to generalize to different speakers, in bilingual and monolingual infants (Henderson & Scott, 2015; Scott & Henderson, 2013). Finally, bilingual and monolingual infants adapted their reasoning about another person’s knowledge based on whether they heard one language or two being spoken to this person (Colomer & Sebastian-Galles, 2020; Pitts et al., 2015). Based on these studies, we can conclude that starting in the first year of life, in some situations infants make inferences about others based on the language they observe being spoken.

There is also evidence that infants take visual cues about a person into account when reasoning about what language an unfamiliar talker might use. At least by 11 months, monolingual English-exposed infants associated faces of East Asian women with hearing Cantonese (May et al., 2019), indicating that they use the specific language spoken as a cue to a person’s identity in some cases. Further studies also show that in certain situations, infants take the speaker’s race into account in language processing tasks (Singh et al., 2020; Weatherhead et al., 2021; Weatherhead & White, 2018). These studies further underline that infants pay attention to the social context of language use.

Current Studies

We tested whether bilingual and monolingual infants spontaneously form an association between a person and the language they speak. Based on the studies of monolingual infants using language to reason about other people, as well as infants’ perceptual abilities to discriminate languages and speakers, we expected that all infants would spontaneously form an association between a person and the language they speak. While there is evidence that bilingual adults form expectations about others’ language use (Martin et al., 2016), whether or not bilingual and monolingual infants can learn to associate a person and the language they speak in a controlled lab environment has not been tested. The present paper will address this gap.

We tested infants using auditory-only stimuli in Study 1, and audiovisual stimuli in Study 2. The design of Study 1 was based on the previous literature on language discrimination, where most research uses auditory-only paradigms ((Gasparini et al., 2021). In Study 2, we tested whether the addition of rich multimodal information changes infants’ ability to learn person-language associations. We predicted that infants would have an easier time learning the person-language associations when exposed to audiovisual stimuli (Study 2), as the additional visual clues to the person’s identity and language (Sebastian Galles et al., 2012; Weikum et al., 2007) might make it more engaging and salient to infants. However, we note that in a study reported after our data was collected, audiovisual stimuli possibly made it harder to detect a language switch for close language pairs (Birulés et al., 2018).

Each study used a familiarization-switch procedure. In the familiarization phase, infants encountered a woman speaking English and a man speaking French (or vice-versa). Then, infants were shown “Same” trials with a familiar speaker-language pairing (e.g., the woman speaking English) and “Switch” trials with a new speaker-language pairing (e.g., the woman speaking French). If infants are able to associate a speaker with their language, we expected them to look longer to the Switch than to the Same trial. We designed Study 2 such that we could also investigate pupil dilation as a measure of increased cognitive processing (Sirois & Brisson, 2014). Pupil dilation has been shown to be sensitive to language context (Byers-Heinlein et al., 2017; Fawcett, 2021). We predicted that infants’ pupils would dilate on Switch trials compared to Same trials. This change in pupil size would indicate that infants learned the association between speaker and language in the familiarization, and therefore reacted differently to Switch and Same trials.

The age range tested in our studies spanned from 5 to 18 months. The age groups chosen represent three different developmental stages: before tuning into the sound structure of their native language (5 months), the early word learning phase (12 months), and the so-called vocabulary spurt phase (18 months). In Study 1, we collected full samples for two age groups: 5- and 12-month-olds. Based on the results of Study 1, in Study 2 we tested 12- and 18-month-olds, as we thought older infants might be more successful in our task. We expected that older infants would outperform younger infants, but we were also interested in seeing when infants would first succeed in our task.

We investigated the role of infants’ language background by testing bilingual and monolingual infants. Bilinguals were learners of French and English, so they were familiar with both languages used in this study. Monolinguals had regular exposure to either English or French, and thus were familiar with only one of the two languages used. Our predictions for the role of language background were tentative. On the one hand, bilinguals might outperform monolinguals, as they are familiar with both languages and have regular opportunities to encode speaker-language pairings in their everyday life. On the other hand, bilingual infants might be used to a speaker switching to a new language, and in this case may not respond differently to Switch and Same trials. Language switches may be more obvious to detect for monolingual infants, for whom a speaker switching into or out of their native language is very salient as they hear few language switches in their daily life. Under this assumption, monolingual infants would perform better than bilingual infants.

Results from both studies showed that making person-language associations in a simple lab task is difficult for infants, and that the explanation of how bilingual infants learn to discriminate and make sense of their social environment is potentially more complicated than expected.

Study 1: Auditory Cues Only

Method

This study was conducted in accordance with the Declaration of Helsinki, and was approved by the Human Research Ethics Board of Concordia University (certificate no: #10000439). Parents were informed of the study’s purpose and procedure and provided written consent prior to participation.

Participants

A total of 84 infants aged 5 or 12 months were included in the final sample. Each infant contributed data at a single age. Infants were recruited from government birth lists and a database of interested families in Montreal, Canada, a French-English bilingual city. Data collection took place between November 2013 and February 2017. When planning the study, we aimed for a sample of 16 per age and language group, which was a typical sample size at the time the study was planned (Oakes, 2017). We conducted a sensitivity analysis using G*Power (Faul et al., 2007) for a repeated measures ANOVA with a within-between interaction. The design has a power of 80% to detect an effect of at least d = 0.51. Due to a delay between data collection and data analysis, we tested more than 16 infants in all groups, and included all infants who were eligible based on our exclusion criteria. Infants were determined to be monolingual or bilingual based on their exposure to English and French.

Language exposure was measured via the Language Exposure Questionnaire (Bosch & Sebastian Galles, 2001) using the Multilingual Approach to Parent Language Estimates (MAPLE, see Measures for more detail; Byers-Heinlein et al., 2020). Bilingual participants were exposed to a minimum of 25% English and 25% French, and did not have more than 20% of exposure to a third language. The average exposure of bilingual infants to their dominant language was 62% (range: 43–75%), and to their non-dominant language 36% (range: 26–50%). Nine bilingual infants also had exposure to a third language, at an average of 9% (range: 2–19%). Monolingual participants were exposed to either English or French at least 90% of the time for an average of 98% (range: 92–100%). Bilingual infants’ exposure to language mixing in their daily lives was measured by the Language Mixing Questionnaire (Byers-Heinlein, 2013). Language Mixing Questionnaire was not completed for 5 infants, those were treated as missing data. Bilingual infants scored on average 12.3 out of 30 (range: 0–30), with 30 indicating the highest level of language mixing.

To get to this final sample, 60 additional infants were tested but not included in the analysis due to failure to meet age criteria (n = 3), language criteria (n = 27), low birth weight or premature birth (< 2500 g at birth or < 37 weeks of gestation, n = 3), reported major health issues (n = 1), and experimenter error (n = 1). Data for all remaining infants were analysed to verify if they contributed enough data. For trials to be included into the final analysis, infants had to look for at least one second. Then, for infants to be included in the analysis, they had to have completed at least one test trial of each type (Switch, Same). Following this procedure, in the 5-month-old group we excluded 5 infants (2 bilingual, 3 monolingual) due to insufficient data, and in the 12-month-old group 20 infants were excluded (5 bilingual, 15 monolingual). This resulted in a final analyzed sample of 84 participants across the two age groups. Children’s reported ethnicities were varied: the most reported category was European descent (M = 39 %), followed by parents listing multiple categories (M = 23%), including combinations of Afro-Canadian, Caribbean, and other identities, and some parents used the write-in option to indicate they identified as either Canadian, Quebecois, or North American (M = 14%) or checked unknown (M = 12%). Furthermore, 6% indicated Afro-Canadian ethnicity, and all other options had less than 5% responses. Families had a high average level of maternal education (M = 15.9 years). A full breakdown of participants by age group and language group is displayed in Table 1.

Table 1. Descriptive statistics for child participants by age group and language background.

			Age (months)
Age	Language group	N	Min	Mean	Max.	% Female	% English dominant
Study 1: Auditory-only
5m	Bilinguals	19	5.0	5.5	6.0	63	53
	Monolinguals	19	4.8	5.7	6.5	42	42
12m	Bilinguals	24	11.7	12.3	13.0	54	67
	Monolinguals	22	11.8	12.5	3.3	36	41
Study 2: Audiovisual
12m	Bilinguals	18	11.8	12.5	13.2	33	67
	Monolinguals	24	11.8	12.6	13.2	42	67
18m	Bilinguals	21	17.7	18.5	19.2	67	52
	Monolinguals	18	17.8	18.5	19.1	50	56

Note. For monolinguals, the column % English dominant denotes the percent of children who are English monolinguals, the remainder of monolinguals are French-dominant.

We had also planned to test 9- and 18-month-old infants, and bilingual infants who hear either English or French as well as another language (“bilingual-other”), however due to recruitment difficulties these samples were not completed. Data from these infants were not analyzed for the purposes of this paper. For transparency, we have shared more details on these incomplete samples in the Supplemental Materials and the full dataset is available on the OSF (https://osf.io/psbxv/).

Stimuli

Auditory stimuli consisted of voice recordings made by one male and one female English-French bilingual speaker. Both speakers had acquired Canadian English and Quebec French from birth and spoke with a Montreal-native accent in either language as judged by a different highly proficient English-French bilingual. The passages that were recorded for the current study were excerpts from The Little Prince, which was chosen as it was used in Weikum et al. (2007). The speakers read the passages in adult-directed speech, as most of the studies on language discrimination have used adult-directed speech (Gasparini et al., 2021). For each trial, we selected sentences read by a single speaker, such that pitch and duration were matched across speakers and languages. All recordings were normalized to 70dB using PRAAT (Boersma & Weenink, 2016) and each trial lasted approximately 18 seconds. A transcript of the passages used and stimuli files can be accessed on the OSF platform (https://osf.io/psbxv/). The visual stimulus was constant throughout the study, and consisted of a picture of a field of flowers (see Fig. 1) which was chosen to provide a visually pleasing stimulus for infants to look at.

Figure 1.

Example of one Trial Order for Studies 1 and 2

Note: This figure shows two of four test trials to illustrate, in the remaining test trials the woman speaks either the same or a different language. Trial order was pseudorandomized, and assignment of speaker and language combination was counterbalanced across infants. The outline colour of the speech bubble represents the language spoken, English (red) or French (blue).

Measures

Demographic information.

Parents completed a demographics questionnaire that asked about infants’ health history, gestational age, birth weight, family structure, and parents’ level of education.

Multilingual Approach to Parent Language Estimates (MAPLE).

An adapted version of the Language Exposure Questionnaire (Bosch & Sebastian Galles, 2001) was administered as a structured interview using MAPLE (Byers-Heinlein et al., 2020). The experimenter asked questions about language exposure on a typical day for each month in the child’s life and then calculated the percentage of the child’s exposure to English, French, and other language(s). This information was used to calculate the percentage of time that infants had been exposed to each language since birth.

MacArthur-Bates Communicative Development Inventory (MCDI).

In order to assess infants’ vocabulary size for comprehension and production, parents of children 12 months or older filled out either the MCDI – Words and Gestures (12-month group) or MCDI – Words and Sentences (18-month group). Parents of monolingual infants completed the MCDI in its original English version (Fenson et al., 2007) or in its Canadian French in Quebec adaptation (Trudeau et al., 1999) as applicable, while parents of bilingual infants completed both versions.

Language Mixing Questionnaire.

Parents completed a self-report questionnaire that assessed the frequency of language mixing when speaking directly to their child, including code switching and borrowing (Byers-Heinlein, 2013). This measure was only collected for bilingual infants.

Procedure

Participants were tested in a dimly lit sound-attenuated room. Infants sat on their parent’s lap facing the screen of a Tobii T60-XL eye tracker (Tobii^® Technology, 2010) that gathered eye-gaze data. Parents wore darkened sunglasses and listened to music on headphones to minimize their interaction with the infants during the study. The experimenter controlled the study from an adjacent room using Tobii Studio software and was blind to the trial type presented. Before stimuli presentation, the eye-tracker was calibrated to the participants’ eyes using a five-point infant calibration routine. A bouncing circle changing in colour and shape was presented before each trial to redirect the infant’s gaze to the centre of the screen.

Infants were familiarized to a total of 8 trials of either the male or the female speaker talking, interleaved in a pseudorandomized order. Each trial used a different passage from the story The Little Prince. Infants heard a consistent speaker-language pairing, e.g., the female voice speaking English and the male voice speaking French. The speaker-language pairing was counterbalanced across infants, such that some infants heard the female speaker speaking French and the male speaker speaking English and vice versa. In the test phase, infants heard two trials with the same speaker-language pairing as in familiarization (Same trials) and two trials with each of the familiarization speakers talking in a different language (Switch trials). All infants saw four test trials, one of each trial type per speaker. The order of test trials was also counterbalanced, such that some infants heard a Switch trial first and some heard a Same trial, and some heard the male speaker first and some heard the female speaker first. An example of a study order is shown in Figure 1. The eight familiarization trials within each order were presented quasi-randomly in one of eight experimental orders, such that no more than two trials of the same speaker-language pairing appeared in a row. The experiment lasted approximately 5 minutes. After completion of the study, infants received a participation certificate and a small gift.

Results

Analytic Approach

The area of interest was defined as the entire display area (1920×1200). The dependent variable was total looking time in seconds on each test trial, which we log-transformed following recommendations for infant looking time studies (Csibra et al., 2016). As a robustness check we repeated all looking time analyses with the untransformed looking times, and report the full results tables in the Supplemental Materials as well as a short summary in the sections below. As described in the Participants section, infants who looked for less than 1 s on more than two of the four test trials were excluded from analysis, and following that, infants who did not contribute at least one type of each trial (Switch, Same) were excluded. The remaining 5-month-olds completed on average 12.0 trials with 6.41s of looking time per trial, and 12-month-olds completed on average 11.9 trials with an average looking time of 4.01s per trial. As a check, we examined looking times over the course of familiarization trials, which confirmed that infants’ attention decreased significantly from the first familiarization trial (M = 10.1s, SD = 3.9s) to the final familiarization trial (M = 5.8s, SD = 4.0s, t(83) = 9.49, p < 0.001, d = 1.09). Further exploratory analyses of correlations of infants’ performance in our task with language variables can be found in the Supplemental Materials.

ANOVA and Equivalence Tests

A 2 (language background: monolingual, bilingual) × 2 (test trial type: Same, Switch) mixed factorial ANOVA was conducted separately for each age group. Language background was a between-subjects factor and trial type was a within-subjects factor. The main dependent variable was log-transformed looking times, and note that the same pattern of results was found in the robustness analysis (Table S1 in the Supplemental Materials). Mean looking times can be seen in Table 2, difference scores of looking time in Switch and Same trials are visualized in Figure 2 (left panel).

Table 2. Mean looking times by Study, Age and Language Group.

		Looking Times (s)
		Same		Switch
Age Group	Language Group	M	SD	M	SD	Cohen’s d
Study 1: Auditory-only
5-month-olds	Bilinguals	6.03	2.81	6.08	3.07	0.01
	Monolinguals	6.79	3.66	6.76	3.54	-0.01
12-month-olds	Bilinguals	4.53	2.87	3.90	2.35	-0.24
	Monolinguals	3.70	2.24	3.86	2.14	0.07
Study 2: Audiovisual
12-month-olds	Bilinguals	7.49	3.47	6.49	3.41	-0.29
	Monolinguals	8.98	3.44	8.40	3.52	-0.17
18-month-olds	Bilinguals	9.48	5.52	10.87	4.65	0.27
	Monolinguals	8.25	5.04	8.38	4.27	0.03

Figure 2.

Difference scores for looking to Switch versus Same trials by study and age group. Note. Black dots with whiskers inside the boxplots show the mean and the standard error of the mean. Violin plots show the distribution of values.

For the 5-month-olds, we found no effect of trial type (F[1, 36] < 0.01, MSE = 0.13, p = .955, η²p < .01), indicating that infants did not look differently to Switch trials (M = 6.42) than to Same (M = 6.41, Cohen’s d < 0.01). There was no effect of language background or interaction between language background and trial type (language background: F[1, 36] = 0.26, MSE = 0.57, p = .615, η²p = .007; interaction: F[1, 36] < 0.01, MSE = 0.13, p = .981, η²p < .01) in the 5-month old group.

Similarly, for the 12-month-olds, we also found no effect of trial type (F[1, 44] = 0.04, MSE = 0.26, p = .842, η²p = .001), indicating no difference in looking time to Switch (M = 3.88) and Same trials (M = 4.13, Cohen’s d = − 0.10). There was also no effect of language background or interaction with language background (language background: F[1, 44] = 0.36, MSE = 0.57, p = .554, η²p = .008; interaction: F[1, 44] = 1.04, MSE = 0.26, p = .313, η²p = .023).

To summarize, across 5- and 12-month-old infants, there was no evidence that infants noticed when a speaker switched languages. Infants’ prior language exposure did not seem to affect their behaviour.

We used equivalence tests to test if there may be a true effect of trial type that our analysis was underpowered to detect. Unlike null-hypothesis tests, which can only conclude whether two conditions are significantly different, equivalence tests allow us to conclude that two conditions are statistically equivalent. Equivalence tests necessitate choosing the smallest effect size of interest, which we set to be the smallest effect size we can detect with our planned sample size based on our sensitivity test, d = 0.51. Equivalence tests indicated that proportional looking time to Switch and Same trials was statistically equivalent, for both 5-month-olds (t[37] = 3.10, p = 0.002) and 12-month-olds (t[45] = 3.27, p = 0.001). The same pattern of results was found in the robustness analysis using untransformed looking times, and can be seen in Table S2. This indicates that we can reject the hypothesis that the true effect size is larger than d = 0.51 (or smaller than d = −0.51). We thus conclude that infants’ equal looking times on auditory-only Same and Switch trials are likely a true null result.

Discussion

In Study 1, bilingual and monolingual infants aged 5- and 12-months-old did not show evidence of forming person-language associations. One reason why infants may not have formed these language-person associations could be because they did not have access to the facial cues associated with seeing someone speak that they would have in real world interactions. These cues have been shown to allow infants to detect a language switch in a visual-only task (Sebastian Galles et al., 2012; Weikum et al., 2007). Furthermore, adding visual information might make the speaker’s identity more salient and might therefore facilitate the infants’ learning about the speaker. Thus, Study 2 examined whether infants could form person-language associations when both seeing and hearing a speaker. In addition to measuring looking times, we also designed this study such that we could measure pupil dilation, by having the onset of speech start after the onset of the visual stimulus, to be able to measure any effects of cognitive processing in pupil dilation separately from the luminance changes associated with the onset of the visual stimulus. Together, looking times and pupil dilation allow us to test whether infants learned the association between a person and the language they speak in an audio-visual task.

Study 2: Auditory and Visual Cues

Method

The procedures for Studies 1 and 2 were identical, except for the video stimuli used in Study 2. Differences in protocols are outlined below.

Participants

A total of 81 bilingual or monolingual infants were included in the final sample. We recruited a 12-month-old group and an 18-month-old group; 5-month-olds were not tested as this population was harder to recruit than slightly older infants and did not show significant effects in Study 1. All other criteria for participation were identical to Study 1. Data collection took place between September 2017 and September 2019. We planned to test 16 infants per group to match the samples for Study 1, and for all groups we ended up with a slightly larger group of participants due to delays between recruitment and data analysis, as in Study 1. Twenty-seven additional infants were tested but not included in the analysis due to failure to meet age criteria (n = 1), pre-established language criteria (n = 20), or low birth weight or premature birth (n = 6). As in Study 1, data for the remaining infants were analysed to check if they contributed enough data. Following this procedure, in the 12-month-old group four infants were excluded (4 monolinguals), and in the 18-month-old group four infants were excluded (1 bilingual, 3 monolingual). This resulted in a final analyzed sample of 81 infants across the two age groups. More information on the samples can be seen in Table 1.

As in Study 1, language exposure was measured using a structured interview (Bosch & Sebastian Galles, 2001; Byers-Heinlein et al., 2020). Bilingual infants’ average exposure to their dominant language was 62% (range: 44–75%), and to their non-dominant language 35% (range: 25–49%). Nine bilinguals had some exposure to a third language, on average that exposure was 13% (range: 2–19%). Monolingual infants were exposed to an average of 98% of either English or French (range: 90–100%). Parental responses on the Language Mixing Questionnaire (Byers-Heinlein, 2013) were collected for bilingual infants, as in Study 1. Bilingual infants scored on average 10.0 out of 30 (range: 1–24), with 30 indicating the highest level of language mixing (for one infant, the questionnaire was not completed, this data point was treated as missing).

The reported ethnicities of our sample were as follows: most reported European descent (51%), followed by multiple responses (33%), including Afro-Canadian and Caribbean. All other categories had less than 5% responses. Maternal education was high, with on average 16.4 years of education.

Stimuli

Audiovisual stimuli consisted of video recordings of one male and one female English-French bilingual speaker reading the same passages used in Study 1. As the two original speakers recorded in Study 1 were not available, we recorded two different early learners of Canadian English and Quebec French. Each speaker was filmed in front of a white background while wearing a black shirt. Speakers were instructed to keep a neutral face and to move their head as little as possible. Each video started with ~2 s of silence before speech onset, where the speaker looked straight at the camera. This was necessary for analysing the pupil dilation at the speech onset, to avoid conflating the onset of the video and the corresponding luminance changes with the infant’s processing of the onset of speech. We used the same passages from The Little Prince as in Study 1. The passages were read in adult-directed speech and each one lasted approximately 20 s. As in Study 1, each video clip consisted of a single speaker reading a single passage. The audio of all recordings was normalized to 70dB using PRAAT and the videos’ white balance was adjusted using iMovie. An example of the audiovisual stimuli is shown in Figure 1, and all stimuli files are also available on the OSF platform (https://osf.io/psbxv/).

Measures and Procedure

Parents completed the same questionnaires as in Study 1. The same familiarization-test paradigm and design as in Study 1 was used.

Results

Analytic Approach

The analytic approach for looking time data was parallel to Study 1. Once again, our main analysis focused on log-transformed looking times, and we conducted robustness analyses on raw looking times. Additionally, we were also able to analyze pupil dilation data. This was possible in this study but not in Study 1 because we designed the stimuli such that the speech onset (which differentiates Switch and Same trials) started after the infants’ pupils had adjusted to the change in luminance associated with the onset of the video stimulus. Twelve-month-olds on average completed 11.6 trials with a mean looking time of 7.96 s per trial, and 18-month-olds completed 11.5 trials and had a mean looking time of 9.32 s per trial. As was done for Study 1, we examined looking times over the course of familiarization trials, which again confirmed that infants’ attention decreased significantly from the first familiarization trial (M = 15.7s, SD = 3.8s) to the final familiarization trial (M = 10.2s, SD = 5.2s, t(80) = 8.39, p < 0.001, d = 1.22). Further exploratory analyses of correlations of infants’ performance in our task with language variables (same as for Study 1) can be found in the Supplemental Materials.

ANOVA and Equivalence Tests

Mean looking times can be seen in Table 2, difference scores of looking time in Switch and Same trials are visualized in Figure 2 (right panel).

As in Study 1, for each age group, we calculated separate 2 × 2 ANOVAs with trial type (Same, Switch) as a within-subjects factor and language background (bilingual, monolingual) as a between-subjects factor. In the 12-month-old group, effect of trial type trended towards but did not reach statistical significance (F[1, 40] = 3.89, MSE = 0.09, p = .056, η²p = .089). Twelve-month-olds looked numerically shorter looking to Switch trials (M = 7.58) than Same (M = 8.34, Cohen’s d = −0.22). The effect of language group trended towards but did not reach statistical significance (F[1, 40] = 3.53, MSE = 0.39, p = .067, η²p = .081). In the robustness analysis on untransformed looking time data, this effect was more clearly not statistically significant (p = .124, η²p = .058, see also Table S1). Numerically, bilingual infants (M = 6.99s) looked less long than monolingual infants (M = 8.69s, Cohen’s d = −0.49), regardless of trial type. There was no interaction between language group and trial type (F[1, 40] = 0.48, MSE = 0.09, p = .491, η²p < .012). The pattern of results in the robustness analysis was the same, except where indicated otherwise in the text above.

In the 18-month-old group, we found no effect of trial type (F[1, 37] = 1.76, MSE = 0.24, p = .193, η²p = .045), as evidenced by similar looking times on Switch (M = 9.72s) and Same trials (M = 8.91s, Cohen’s d = 0.16). Again, there were also no effects of language group (F[1, 37] = 1.18, MSE = 0.66, p = .284, η²p = .031) or interaction between language group and trial type (F[1, 37] = 0.49, MSE = 0.24, p = .488, η²p = .013). The pattern of results in the robustness analysis on untransformed looking time data was the same. While we did find a marginal difference between Switch and Same trials in the 12-month-olds, the direction was opposite to the expected effect, and different to the direction of the effect in 18-month-olds. We thus speculate that the marginal effect found for 12-month-old is likely due to chance.

We used equivalence tests to determine whether infants’ looking times for Switch and Same trials were statistically equivalent. As in Study 1, we used d = 0.51 as our smallest effect size of interest, which is the smallest effect size our planned sample has 80% power to detect. In the 12-month-old group, the equivalence tests were not significant (t[41] = −1.33, p = 0.095), which means we cannot reject the hypothesis that the true effect is larger than d = 0.51 (or smaller than d = –0.51). However, in the robustness analysis on untransformed looking time, the equivalence test was significant (t[41] = −1.73, p = 0.046, see also Table S2 in Supplemental Materials). These conflicting findings in conjunction with the marginal but not statistically significant main effect of trial type in the ANOVA, indicate that our observed data are unsurprising in either case, whether they were sampled from a distribution around 0 (as tested in the ANOVA) or a distribution around d = 0.51 (as tested in the equivalence test). For the 18-month-olds, the equivalence tests were significant (t[38] = 1.86, p = 0.035), meaning that for 18-month-olds we can reject the hypothesis that the effect size is larger than d = 0.51 (or smaller than d = −0.51). The results of the robustness analysis on untransformed data showed the same result (see Table S2). In summary, there may be a smaller-than-expected effect in the opposite direction than predicted in 12-month-olds, which would need to be explored in future research with a larger sample size. In the 18-month-olds sample, we found equal looking times to Switch and Same trials based on a smallest meaningful effect size of d = 0.51. Taken together, there is no evidence that 18-month-olds are affected by whether the speaker switches into or out of their dominant language when observing a person switching languages, and inconclusive evidence for 12-month-olds, which necessitates future research with a larger sample size.

Pupil Dilation

We investigated pupil dilation as a measure of increased cognitive processing in response to Switch compared to Same trials. We followed recommendations for analyzing pupillometry data in our data preparation (Forbes, 2020; Sirois, 2018) . In cases where pupil data was missing for only one eye, we regressed the values from the available eye, since pupil dilation measures for both eyes were highly correlated (r = .94). We then averaged pupil size across both eyes, and down sampled the data to 200ms time bins. We removed trials with less than 25% of valid pupil size measurements, which was 95 trials (31.9% of all trials). We then removed infants who did not contribute at least one trial of each type (Same, Switch), which removed 20 participants (24.7% of the sample). This resulted in 31 12-month-olds (12 bilinguals, 19 monolinguals) and 30 18-month-olds (16 bilinguals, 14 monolinguals). We then filtered data using a median filter to remove noise, and interpolated across missing data such as blinks using linear interpolation. Finally, we set the baseline to the 200ms time window before onset of speech, and subtracted the mean pupil size during the baseline from the within-trial pupil sizes to get change in pupil size from baseline.

We then computed separate 2 × 2 ANOVAs for each age group, with pupil dilation as dependent variable, and trial type (Switch, Same) as a within-subjects variable and language background (bilingual, monolingual) as between-subjects variable. In 12-month-olds, we found no effect of trial type (F[1, 29] = 0.01, MSE < 0.01, p = .909, η²p < .001), as evidenced also by similar change in pupil size for Switch (M = 0.085) and Same trials (M = 0.084) . We also did not find an effect of language background (F[1, 29] = 0.02, MSE = 0.02, p = .894, η²p = .001) or an interaction between language background and trial type (F[1, 29] = 1.03, MSE < 0.01, p = .317, η²p = .034). In 18-month-olds, we found no effect of trial type (F[1, 28] = 1.13, MSE = 0.01, p = .297, η²p = .039), as evidenced by similar pupil dilation on Switch (M = 0.110) and Same trials (M = 0.088). We further found no main effect of language background (F[1, 28] = 1.85, MSE = 0.01, p = .184, η²p = .062) or interaction between language background and trial type (F[1, 28] = 0.01, MSE = 0.01, p = .934, η²p < .001). Thus, we did not find evidence that infants noticed the speaker switching languages when using pupil dilation as a proxy for increased cognitive processing.

General Discussion

The current studies investigated whether bilingual and monolingual infants keep track of who speaks which language (i.e., spontaneously form a speaker-language association), by testing whether they would detect when speakers switched languages. In Study 1, we tested 5- and 12-month-olds using an auditory-only task and found no evidence that infants made an association between speaker and language spoken. In Study 2, we tested 12- and 18-month-olds and again found no evidence that infants made associations. In both studies, there was no difference between bilingual and monolingual infants. The ability to spontaneously form speaker-language associations relies on several prerequisite abilities, any of which could have contributed to infants’ failure in the current study: (1) discriminating languages, (2) discriminating people, and (3) associating a specific language with a specific person. We will discuss each of these in turn, before turning to the role of language experience and task modality.

First, infants must be able to discriminate between different languages in order to associate different languages with different people. Our study used English and French, which have different rhythmic and stress patterns, cues which are known to support language discrimination (Gasparini et al., 2021) Moreover, numerous studies across ages and paradigms have shown that infants can readily discriminate English and French, for example 2-month-olds in auditory language discrimination tasks (Dehaene-Lambertz & Houston, 1998), 4- to 8-month olds in visual language discrimination tasks (Weikum et al., 2007), as well as looking-while-listening tasks in older infants (e.g., Byers-Heinlein et al., 2017). Although we did not test infants’ ability to discriminate our particular stimuli, based on these past studies, including some which used similar stimuli (Weikum et al., 2007), we expect that infants were able to discriminate between English and French in our study and that their failure on our task was not due to an issue with language discrimination.

Second, infants must be able to discriminate different individuals in order to associate them with a language. Our study used one male and one female speaker, given that male and female voices are discriminated early in development (Johnson et al., 2011; Miller, 1983). In Study 2, infants heard and saw the speakers, making it even less likely that they did not distinguish between speakers. We think it is therefore unlikely, but not impossible, that infants failed to discriminate between the two speakers.

Third, it could be that infants discriminated languages and the speakers, but that they struggled to make the association between a speaker and the language spoken. The ability to make associations is crucial across many aspects of learning, including learning about the sounds, words, and rules of a language (Saffran & Kirkham, 2018). However, infants make some types of associations more readily than others. For example, infants can learn that some speakers systematically pronounce some words differently, and generalize that to new situations (Weatherhead & White, 2016). Moreover, even within the same apparent task such as word-object associations, infants succeed in some paradigms (for review, see Tsui et al., 2019), but not in others (Gonzalez-Barrero et al., 2021). It may be that speaker-language associations are relatively difficult to form, or are better formed under other learning circumstances than the one we tested here. Tangentially related here is some evidence that 26- to 28-month-old bilingual children interacting with strangers modulate their language use somewhat according to the strangers’ language proficiency (Comeau et al., 2003; Genesee et al., 1996), however those studies observed a small sample of 4–8 children each and thus it is difficult to draw firm conclusions. Overall, infants might require a high degree of familiarity with speakers before they form a speaker-language association, and therefore might do so with well-known speakers in everyday life, but not spontaneously in a short laboratory task.

Our study also explored two potential moderators of infants’ performance: the infants’ language experience that they bring to the task (bilingual vs. monolingual), and the modality of the task. Our results did not show a difference for either of these two factors.

With regards to language experience, we had expected to find that both bilingual and monolingual infants would form person-language associations, though possibly through different mechanisms and at different ages. Bilingual infants are exposed to two languages regularly, and we therefore expected them to pay attention to language switches and possibly to language-person associations. We expected monolinguals, on the other hand, to notice when a speaker switches out of or into their native language. The population of monolinguals we tested grows up in a bilingual city, however, and thus may be more used to hearing other languages than those living in a monolingual context. Thus, they may not have been surprised to hear a speaker switch into another language. Testing monolinguals living in a more monolingual context could rule out this possibility. We, however, found that neither group appeared to learn the person-language associations in our task.

With regards to modality, we expected infants to have an easier time learning the association in the audiovisual task, which provided more cues to the speaker’s identity (the characteristics of both their voice and their face), and to the properties of the language (both in the auditory modality and information on the face and lips) than the auditory-only task. Indeed, it is hypothesized that redundant information from auditory and visual modalities is important for infants’ development (Bahrick et al., 2004). For example, both auditory and visual cues contribute to performance in speech-in-noise tasks (Lalonde & Werner, 2019) and language discrimination (Sebastian Galles et al., 2012; Weikum et al., 2007). Further, a number of studies have shown that infants and children attend to the mouth of a speaker (e.g., Lewkowicz & Tift, 2012; Morin-Lessard et al., 2019), presumably to aid comprehension. Thus, the audiovisual condition provided infants with more information overall, and was potentially more engaging. However, we found similar patterns of results for the auditory-only and audiovisual studies, as infants did not learn the speaker-language association in either study. This result is not unprecedented in the literature. In a study of language discrimination, Birulés et al. (2018) reported that bilingual 4-month-olds, when given audiovisual cues, did not discriminate Catalan and Spanish, a language pair that other studies have shown to be discriminable with auditory-only cues at the same age (Bosch & Sebastian Galles, 2001). Another study found that in some situations, seeing a speaker’s mouth did not improve performance in a word recognition task (Singh et al., 2021). Together with the results of our study, this indicates that more information may not always result in better performance in infant learning (see also Newport, 1990).

Our research question also intersected with the rich literature on monolingual and bilingual infants’ use of language spoken as a social clue. At first glance, the lack of evidence for infants’ ability to associate a person with a language seems contradictory with the findings that infants change their behaviour when observing people speak a certain language. However, there are several differences between our and their studies to consider. First, many of the previous studies tested preference for people speaking a certain language (e.g., Buttelmann et al., 2013; Kinzler et al., 2007) rather than infants learning of and recalling associations between a person and a language. In our task, infants had to learn the association between a speaker and a language, and then during the test trials notice that the speaker switched to a different language, a potentially demanding task. Second, other studies have relied on experiences that infants have made in their daily lives, such as about the connection between race and language (May et al., 2019; Weatherhead & White, 2018). For example, 3-year-olds make assumptions about the language individuals speak when shown pictures of people of different races, but not when shown pictures of old and young people (Hirschfeld & Gelman, 1997). This connection between race and language, learned from everyday experience, is likely stronger than associations learned in the lab. In our study, people of different genders spoke different languages, which may not be a typical regularity in the outside world. For regularities that are typically consistent, such as voice and visual cues to gender, these associations are learned in the first year of life (Hillairet de Boisferon et al., 2015; Walker-Andrews et al., 1991). We chose to use a male and a female speaker to maximize the differences between each speaker, however future studies could test for example the infants’ own parents to tap into infants’ knowledge of the language preferences of the people around them (see also Molnar & Davidson, 2017). Third, infants might readily track information about who speaks which language when viewing multi-speaker interactions (Colomer & Sebastian-Galles, 2020; Pitts et al., 2015; Vouloumanos, 2018), but not in the single-person videos our study used. Finally, it may be that infants are able to make language-based inferences, (e.g., whether speakers of different languages will know the same object label, Henderson & Scott, 2015; Scott & Henderson, 2013) without directly forming speaker-language associations, for example by reasoning about group membership. These discrepancies between our studies’ results and previous studies also highlight the importance of integrating findings across sub-disciplines of cognitive and social child development.

Implications and Future Directions

Our results have several implications for understanding bilingual language acquisition. First, a popular idea in bilingual parenting is that the one-person-one-language approach is the best way to support language acquisition (Barron-Hauwaert, 2004). One assumed benefit of this approach is that it helps infants to discriminate and separate their languages, as they associate different speakers with different languages. However, our results suggest that infants do not seem to readily track who speaks what language, although our study does not address whether infants might form such associations over the long term with their primary caregivers. Indeed, our results are consistent with the general finding that bilingual infants learn from a multitude of family language configurations providing they have sufficient high-quality experiences in each language.

Second, it is possible that the way caregivers switch between languages already provides infants with helpful cues to detecting the switch. For example, studies of language mixing in parents show that parents mostly switch languages between sentences rather than within sentences (Kremin et al., 2022), which may help infants’ comprehension (Byers-Heinlein et al., 2017). Given that we did not find statistically significant results in our task, there is an open opportunity for future research. One avenue would be to use a different paradigm that does not rely on infants being surprised that a person is switching languages, as this may be a common occurrence in infants’ lives. Instead, sensitivity to language switching could be tested by introducing two speakers who each speak a different language, as in our study, and then showing pictures of both speakers while infants hear one language. If infants look more at the person previously associated with the language they are hearing, this would indicate that they made the association between the language and the speaker.

Third, it might be that the differences between languages in certain tasks are less salient than we would assume based on language discrimination studies. A recent study found that infants did not discriminate between English and French when hearing single words that were either in a single language (dog… milk) or alternated between languages (dog... lait [fr. milk], Schott et al., 2021). Whereas in a looking while listening task, 20-months-olds did notice when the language switched between English and French (e.g., Byers-Heinlein et al., 2017). The study on single-word language discrimination used an infant-directed speech comprehension task, whereas our study was modelled after language discrimination studies and therefore used adult-directed stimuli that were not designed to be understood by infants. It is possible that language discrimination happens as a “by-product” of higher-level processes like lexical access of words that are organized by language, which we would not have tapped into using our task, as our task was not designed to require lexical access. Our results suggest that language discrimination is more nuanced than previously thought based on early language discrimination studies.

Finally, it is possible that with tweaks to our study design, infants might have shown an ability to form person-language associations. For example, we used a familiarization paradigm with a fixed number of trials prior to test, but it is possible that a habituation paradigm might have been more sensitive. We note that anecdotally, attrition rates were higher for monolinguals than bilinguals, perhaps due to their unfamiliarity with one of the languages, which might have been ameliorated in an infant-controlled habituation design. It is also possible that a paradigm using live actors rather than recordings might have been more revealing of infants’ underlying abilities. Future studies could investigate these possibilities.

Conclusion

In two studies, we did not find evidence that infants spontaneously formed associations between a person and the language they speak. Although we cannot rule out that this was due to the particularities of our task, this finding hints at the potentially long developmental path for bilingual infants to move from simple discrimination of their languages to more complex forms of language discrimination, separation, and social use (Byers-Heinlein, 2014). This work highlights a need to integrate perceptual and social approaches, to uncover the mechanisms at play in early bilingual language learning.