In face-to-face conversations, we can tell who is speaking by looking at whose mouth is moving. Given the intersensory redundancy between this visual information and the resulting vocal signals, it is tempting to dismiss voice recognition as a trivial skill. However, there are often situations in which a listener may not have access to visual information—for example, during auditory-only telecommunication, when facing away from the speaker, or when the speaker is wearing a face mask. Moreover, familiarity with a speaker’s voice has consequences for social communication and linguistic processing (see Creel and Bregman, 2011). Thus, examining how listeners recognize voices when visual facial information is partially occluded is important for our understanding of how humans process speech. In the current study, we approached this question by examining whether infants can detect a speaker change when the speaker’s face is partially occluded.

From an early age, humans are surprisingly adept at tracking highly familiar voices. Previous studies have reported that upon hearing their mother’s voice (and relative to hearing an unfamiliar woman’s voice), fetuses’ heart rate increased (Hepper et al., 1993; Kisilevsky et al., 2003), newborns preferentially sucked on a pacifier in a non-nutritive sucking procedure (DeCasper and Fifer, 1980), and 4-month-olds looked toward their mother (live: Spelke and Owsley, 1979; photograph: Orena and Werker, 2021). Infants can also differentiate between their father’s voice and an unfamiliar male voice (DeCasper and Fifer 1980; Ward and Cooper, 1999). These studies show that, in the absence of synchronous visual information, infants can match highly familiar voices to the identities of familiar individuals.

A related question is whether infants can discriminate and learn to recognize unfamiliar voices with the same ease. Indeed, processing voices from unfamiliar talkers is a separate and often more challenging task than processing highly familiar voices (Stevenage, 2017). Infants show more attentive and mature processing of speech spoken by highly familiar voices versus an unfamiliar speaker (Purhonen et al., 2004; Naoi et al., 2012). Unsurprisingly, with increased exposure to a certain set of speakers, adult listeners similarly improve at differentiating the voices of those speakers (e.g., Levi et al., 2011).

Research-to-date indicates that infants are highly adept at discriminating between unfamiliar voices, particularly when the pair of voices are highly distinct from each other (e.g., Floccia et al., 2000). Moreover, infants can pair male and female voices with male and female faces, respectively, by 8 months of age (Patterson and Werker 2003). Infants can also successfully discriminate between voices of same-gender pairs of speakers. In a series of studies, researchers reported that, after being habituated to an unfamiliar female voice, infants as young as 4 months of age dishabituated to a new unfamiliar female voice (Johnson et al., 2011; Fecher and Johnson; 2018, 2019).

Of note, much of the infant work on processing and learning unfamiliar voices has been limited to tests of discrimination, rather than recognition. A recent study by Fecher et al. (2019) revealed that even 16.5-month-old infants have difficulty in learning the voices of unfamiliar speakers. In their task, infants were shown pairings of two voices and their identities (either cartoon characters or talking human faces). When the pair of voices involved one male and one female speaker, infants showed learning of the two voices in a preferential looking procedure experiment. However, when the pair of voices involved two female speakers, infants showed no recognition of either voice, suggesting that learning to recognize unfamiliar face-voice pairings may be a challenging task.

Like adults, certain factors appear to modulate infants’ talker processing abilities. For instance, it is now well-established that listeners are better at learning others’ voices when they are speaking a familiar versus an unfamiliar language (Goggin et al., 1991). Some studies indicate that this effect is a function of phonological processing (Perrachione et al., 2011; Kadam et al., 2016). Yet, even long-term systematic exposure to a language appears to facilitate talker processing (Orena et al., 2015), suggesting that benefits to talker processing could emerge prior to comprehension of the language. Indeed, even infants show a language-familiarity benefit to voice discrimination (Johnson et al., 2011). Recently, Fecher et al. (2019) found that 4-month-olds could discriminate between female voices speaking a familiar language, but not when speaking an unfamiliar language. These findings show that phonetic and indexical information is integrated early in speech processing.

In the current study, we examined the nature of early talker processing skills by tackling two research questions. First, we examined the voice recognition abilities of young children. We followed up on work by Fecher et al. (2019) to investigate whether young children will show voice recognition of unfamiliar voices when cognitive demands are eased. In Fecher et al. (2019), infants learned the face−voice pairings during a six-trial training phase before being tested on their recognition of the pairings during a two-trial test phase. In our study, infants were taught the face−voice pairings and then tested on their recognition of the face−voice pairings within the same trial. Given that 16.5-month-olds in Fecher et al. (2019) still had difficulty in learning the voices of unfamiliar speakers, we chose a higher age range (i.e., 24-months-old) to examine if it is more stable at this age point. We also tested a young age range (i.e., 12-month-olds) to examine whether young infants would succeed in this modified task. Importantly, in our work, in the test phase the faces were partially occluded.

As a secondary question, we asked whether, like for adults, phonetic content influences the voice discrimination abilities of young children. Work by Andics et al. (2007) found that segmental information contributes to ease of voice discrimination for adults. In their study, listeners heard blocks of various consonant-vowel-consonant words and had to decide whether the word they heard was produced by the same voice as the preceding word or by a different voice. Results indicated that certain segments—particularly segments that were more sonorant (e.g., [m], [s])—helped listeners discriminate voices more than other segments. Here, we examined whether certain phonetic segments in speech are helpful for infants to discriminate between voices.

In the current study, we examined the voice recognition skills of young children. Specifically, we tested infants’ ability to detect a speaker change while speakers’ faces were partially obscured. Infants’ performance in the task revealed at least two important findings.

First, our findings show that stable voice recognition skills—especially for unfamiliar voices—are present by 24 months of age. This older group of infants looked proportionally more at the target voices during the preset WoA for both Same Speaker and Different Speaker trials. These findings suggest that 24-month-old infants were able to learn some aspect of the first speaker’s voice such that: (i) they continued looking at her when they heard her voice again, even after objects blocked infants’ visual access to the speakers’ faces, and (ii) they showed a disambiguation response and looked toward another speaker when they heard another speaker’s voice.

Second, our secondary analyzes reveal that the level of unfamiliar voice learning depends, in part, on the phonetic content of the spoken words—findings that mirror those found with adults in Andics et al. (2007). The 24-month-olds showed a disambiguation response in Different Speaker trials only during trials when speakers were saying words with high talker information (i.e., tokens with sonorous phonemes). These findings confirm that, even for infants, phonetic content affects infants’ ability to learn unfamiliar voices. The results further confirm that the speech processing system that is sensitive to the integration of indexical and linguistic information is in place by late infancy (e.g., Mulak et al., 2017).

Interpreting data from the 12-month-old infants is less straightforward. Regardless of who was speaking during the critical period of the trial, 12-month-old infants continued to look at the first speaker of the trial. Interestingly, 12-month-old infants also did not show signs of recognition for animal sounds. One interpretation for these findings is that unfamiliar voice recognition is a challenging task for young infants. For example, Fecher et al. (2019) found that 16.5-month-old infants were able to learn the voices of two unfamiliar speakers, but only when the acoustic differences between the speakers were large (one male and one female speaker). When there were two speakers of the same gender, 16.5-month-old infants did not show signs of voice learning.

Why might 12-month-old infants have difficulty with the current task? Firstly, the occluding ferns may have disrupted infants’ learning of the unfamiliar voices. Indeed, visual information can help adults encode and retain aspects of a speaker’s voice (Sheffert and Olsen, 2004), and access to synchronized visual information can facilitate infants’ performance in other speech processing tasks (Hollich et al., 2005). A real-world, similar context is the use of face masks, which cover the lower half of a speaker’s face. Some have raised concerns about whether this visual occluder may compromise speech processing (Green et al., 2021). As the current study did not have an extra condition without any visual occluders, further research is needed to address this question.

Alternatively, it is possible that learning to recognize an unfamiliar voice is a challenging task for all 12-month-olds (as in Fecher et al., 2019). Indeed, voice recognition is demonstrably more difficult than voice discrimination (see Perrachione et al., 2011 for a review of talker processing tasks). The discrimination of two auditory tokens can rely on low-level mechanisms, but associating a novel face with a novel voice is more cognitively demanding. Moreover, in the current study, infants heard speakers produce only three instances of one token before being tested on their recognition of that voice. Infants may thus have needed more exposure to the initial speaker, or a less artificial experiment, to show recognition. Indeed, Fecher et al. (2019) found that learning voices is facilitated when the naturalness and social relevance of the task is increased. Nonetheless, the current study suggests that unfamiliar voice recognition is more challenging at 12 months than at 24 months.

It is important to note that the current study was conducted via the online platform, Lookit (Scott and Schulz, 2017). This platform is the first large-scale crowdsourcing platform for conducting online developmental studies. There are multitude benefits of an online platform, including the ability to efficiently run participants, to recruit more diverse participants, and to continue research during laboratory shut-downs, such as during the COVID-19 pandemic. There has been some success in validating the use of these platforms, including findings that experimental conclusions derived via this platform are comparable to those of in-lab studies (Scott et al., 2017; Smith-Flores et al. 2022). Note, however, that effect sizes appear to be smaller for online studies, which may also explain some of our smaller effects. Others have raised concerns about the replicability of online Lookit experiments (Lapidow et al., 2021), potentially due to parental interference. These concerns further temper our interpretation of the 12-month-old data, especially given that they were not successful in identifying the animals from their sounds.

To conclude, this study examined infants’ ability to learn unfamiliar voices. We found that 24-month-old infants were able to learn an unfamiliar voice sufficiently well to detect a voice change when objects blocked visual access to the speaker’s mouths. These findings are highly relevant to the current pandemic and the increased use of face masks. Particularly, these findings reveal that 24-month-old infants can encode indexical properties of an unfamiliar speaker’s voice, and they can access this information even when visual access to the speaker’s mouth is blocked. Certainly, there are important follow-ups to provide firmer conclusions. For example, face masks affect the acoustics of speech production (Llamas et al., 2008), and it would be interesting to investigate whether children’s perception of speech and voices are affected by these alterations. Moreover, one could ask whether exposure to more speakers—including more speakers with masks—might promote learning in this domain. Indeed, prior work has shown that speaker variability promotes learning in the linguistic domain (e.g., Rost and McMurray, 2009; but also see Bergmann and Cristia, 2018). Nonetheless, the successful evidence showed that infants at 24 months could recognize and learn speaker voices even when the face is partially obscured, complements the growing research showing that adults are able to adapt to mask wearing, with equal recognition of both speech and emotional expressions (Trainin and Yeshurun, 2021). Research on these and related topics will help to improve our understanding of how infants make use of the multisensory information around them to adapt to different contexts.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

The studies involving human participants were reviewed and approved by the University of British Columbia, Behavioural Research Ethics Board. Written informed consent to participate in this study was provided by the participants’ caregiver.

AO, AM, and JW contributed to the conception and design of the study. AM created the stimuli set, coded the looking data, with support from other research assistants, and wrote a section of the manuscript. AO and AM set up the study on Lookit. AO performed the statistical analysis, with guidance from JW, and wrote the first draft of the manuscript. All authors contributed to the article and approved the submitted version.

This research was funded by the Social Sciences and Humanities Research Council of Canada (grants 435-2019-0306 and 895-2020-1004) to JW, an NSERC Undergraduate Student Research Award to AM, and a Fonds de Recherche du Québec—Nature et Technologies Postdoctoral Fellowship to AO.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.