Time after time: Voice perception from first impressions to identity recognition

Abstract

When we hear someone speak, we do not just hear ‘a voice’. If the voice is unfamiliar, we form an often complex first impression by inferring various characteristics about the person. If the voice is familiar, at least to some degree, we may be able to recognise and identify the person to whom the voice belongs. Even though first impression formation and identity recognition can thus be seen as being situatied at two opposing ends of a ‘familiarity continuum’, first impressions and identity recognition functionally serve the same purpose: making sense of who another person is. Theories and empirical work examining impression formation and identity perception from voices have, however, developed largely in isolation from one another, with relatively limited cross-talk. In this paper, I will review some recent findings from the literature on first impression formation from unfamiliar voices and voice identity learning and recognition from familiar(ised) voices. I will ask how impression perception and identity perception may interact and interface with one another along this ‘familiarity continuum’ between completely unfamiliar and very familiar voices, trying to bring together these two literatures. Specifically, I will consider what happens to first impressions when we become increasingly familiar with a person, whether first impressions might have an impact on how (well) voices can be learned and recognised, and when and how identity recognition might take over from ad-hoc impression formation.

Introduction

Voices are rich signals that encode a wealth of information (Belin et al., 2004, 2011; Kreiman & Sidtis, 2011): When we hear another person speak in a language familiar to us, we can readily understand what this person is saying. Beyond the linguistic information conveyed via the spoken words, we can also glean insights into a person’s intentions and their evaluation of the speaking situation by listening to how someone is speaking and using their voice. From this paralinguistic information in speaking style and voice quality, we can infer whether someone is joking about what they are saying or whether they are being serious, whether someone evaluates a situation as formal or informal, or whether they are trying to win us over or are trying to close down the interaction.

Finally, voices also give us clues to who a person is. In this paper, I will look in more detail at how we as listeners perceive this kind of person-related information, helping us to make sense of who we are listening to. I will make the point that when we hear and pay attention to another person’s voice, we are unlikely to perceive it simply as ‘a voice’. Instead, we will take away an impression of what kind of person we have heard based on their voice (Kreiman & Sidtis, 2011; Lavan & McGettigan, 2023). If the person whose voice we have heard is entirely unfamiliar to us, we will form a first impression – did we just hear an older, friendly man or a younger, grumpy woman? We will thus evaluate and infer different person characteristics of a person from a voice (‘person perception’). Once we know a person, at least to some degree, we may also be able to recognise and identify this person based on their voice (identity perception: ‘This is my friend Aisha’).

Recent findings on voice impression formation and voice identity recognition

The literature on impression formation has almost exclusively examined how unfamiliar voices are perceived, that is, voices that listeners hear for the first time and know nothing else about. The identity perception literature has, conversely, mainly examined familiar(ised) voices, drawing clear divisions between how identity is processed from familiar and unfamiliar voices (Stevenage, 2018; Van Lancker & Kreiman, 1987). Similarly, the work on (first) impressions is traditionally perhaps more closely associated with questions grounded in evolutionary psychology and bioacoustics, while voice identity studies have more often tackled questions arising from the fields of cognitive psychology, forensic science, and/or speech science. Through locating themselves on different parts of a familiarity continuum and through being studied from different theoretical perspectives, the literatures on (first) impression formation and identity perception have, as a result, evolved largely in isolation from one another. Below, I will review some of the recent findings and questions asked in the work, looking at first impressions and identity perception.

Encountering a voice for the first time: First impressions from unfamiliar voices

Many of the voices we hear every day are unfamiliar to us, and all the voices that we are familiar with now started as being unfamiliar to us initially. Despite this potential unfamiliarity, we nonetheless quickly form impressions of people based solely on their voices. These first impressions are often multi-variate and valenced (Lavan, 2023a; Lavan & McGettigan, 2023). Although these first impressions are often not accurate reflections of a person’s true characteristics (Foo et al., 2022; Skuk et al., 2025; Todorov et al., 2015), first impressions have nonetheless been shown to guide our behaviours, decisions and social interactions. For example, basic voice properties and the impressions formed when hearing a voice for the first time predict election outcomes (Klofstad, 2016; Mileva et al., 2020; Schild et al., 2022; Tigue et al., 2012) and decisions of landlords about which tenants they choose to rent accommodation to (Purnell et al., 1999).

Shared impressions and idiosyncratic contributions to first impressions

When examining these first impressions, many studies have shown that listeners, on average, will agree with one another on whether a voice sounds overall older, trustworthy, educated and beyond, even from minimal exposure to a voice (Lavan, 2023b; McAleer et al., 2014; Mileva & Lavan, 2023; Sorokowski et al., 2023). How can listeners agree on judgments of a person’s voice that is entirely unfamiliar to them? On the one hand, this kind of agreement can arise because some aspects of the first impressions listeners have are to some degree accurate – that is, listeners can glean genuine information about a person’s characteristics from their voice. Characteristics that can be (accurately) perceived from voices can therefore be considered as being ‘apparent’ characteristics in voices. For example, impressions of sex (female vs. male in cisgendered adults) or age (child vs. adult) tend to be highly accurate as voices have clear acoustic markers, such as pitch and formant spacing, that help distinguish between voices belonging to one category versus another. Taking sex as an example, adult male voices tend to be lower in pitch and have distinctive formant spacing, due to adult males having thicker vocal cords and longer vocal tracts (Owren et al., 2007; Titze, 1989). These diagnostic acoustic cues can be perceived from voices, yielding largely accurate perceptions of sex, which in turn lead to agreement among listeners.

Agreement among listeners can also occur even when their evaluations are not accurate. In such cases, listeners still use acoustic cues or other perceptual properties of voices to infer characteristics of a person from short voice snippets. These cues are, however, not directly linked to the true characteristics of the person but instead reflect shared social stereotypes and shared knowledge and heuristics about the world (Irvine & Lavan, 2025; Stolier, Hehman, & Freeman, 2018; Stolier, Hehman, Keller, et al., 2018). For example, for faces, impressions of higher competence are associated with masculinity (Oh et al., 2019). Similarly, voices that are perceived as more masculine are perceived to be more dominant than voices that are perceived to be more feminine, resulting in stereotyped gender differences in dominance perception (e.g. Lavan et al., 2024; Lavan, Mileva, Burton, et al., 2021; Mileva & Lavan, 2023). Since people can agree on who sounds more masculine and that masculine people are dominant/competent, we observe agreement in ratings of competence and dominance. However, there is in this case no robust evidence that the people, whose faces and voices were assessed as being competent or dominant, will actually possess these characteristics (Aronovitch, 1976; Skuk et al., 2025).

In a similar vein to social stereotypes, first impressions are also shaped by halo and over-generalisation effects (Forgas & Laham, 2016; Nisbett & Wilson, 1977; Zebrowitz & Montepare, 2008). These effects describe people’s biases to extend a perceived positive trait to the overall perception of a person. For example, if listeners perceive a person to be trustworthy based on their voice, they would ‘fill in the blanks’ in their impression and assume that the same person also has many more positive (as opposed to negative) characteristics and traits, such as being kind, competent etc. – again, without there being any evidence to support this overarching positive impression given the complete unfamiliarity with the person (e.g. Schirmer et al., 2020).

While first impressions tend to be on average shared among listeners, recent work has made a case that there is nonetheless scope for impressions to be idiosyncratic, following the saying ‘beauty is in the eye of the beholder’. In line with findings from the face perception literature (Hehman et al., 2017; Hönekopp, 2006; Leder et al., 2016; Martinez et al., 2020; Sutherland et al., 2020), we (Lavan & Sutherland, 2024) show evidence of some idiosyncrasy in first impressions from voices. As such, listeners will systematically evaluate the characteristics of voices differently from other people. Specifically, one person may, for example, consistently judge people who speak with a fast speech rate to be competent, while another person may consistently evaluate people who speak more slowly as being competent. The idiosyncratic contributions to first impressions were particularly pronounced for ‘inferred’ characteristics such as trustworthiness, friendliness and competence, which are characteristics that cannot be perceived from voices in light of e.g. informative acoustic cues, leading to accurate percets. For these characteristics, the amount of variation explained in first impressions that is due to idiosyncratic perception is similar to the amount of variation explained associated with shared perceptions, where listeners generally agree with one another. In contrast, for ‘apparent’ characteristics like sex, age and health, individual differences in perception are minimal, given the availability of informative cues in the voice. These findings thus highlight the significant role of individual differences and idiosyncratic perceptions in forming first impressions of inferred characteristics, in particular.

Forming first impressions from minimal auditory input

The existing literature on how first impressions are formed from voices has often also stressed how quickly impression formation appears to occur. This claim can be traced to studies that show that hearing a single word, such as ‘Hello!’, is already sufficient to form a first impression of ‘trait’ characteristics, such as trustworthiness, attractiveness and dominance (Baus et al., 2019; McAleer et al., 2014; McAleer & Belin, 2018; Pernet et al., 2015). Given that it only takes around half a second to say ‘Hello’, this confirms that first impression formation indeed does not require a lot of voice information.

But just how quickly and based on how little voice information can listeners form these very first impressions? I used a perceptual gating paradigm to answer this question, where listeners heard voices, producing steady-state vowels, for different exposure durations (25ms, 50ms, 100ms, 200ms, 400ms and 800ms). Participants then rated 9 characteristics that participants frequently infer when hearing unfamiliar voices (Lavan, 2023a): Sex, age, health, poshness, professionalism, educatedness, attractiveness, trustworthiness and dominance. 25ms of exposure to a vowel (e.g. ‘ah’, ‘ooh’) was enough to form an impression of a person’s sex, age, and health. These ‘apparent’, physical characteristics are to some degree marked acoustically in the voice (see above), such that impressions may be formed based on very short exposure to the relevant diagnostic acoustic features. Surprisingly, impressions of dominance (for male voices only) could also be formed from even minimal exposure. Impressions of other characteristics, such as trustworthiness, educatedness, and professionalism, were formed more slowly and required several hundred milliseconds of exposure to a voice (Lavan, 2023b; Mileva & Lavan, 2023)

This perceptual gating study addressed the research question of how little acoustic information listeners might need to piece together an impression in a perceptual experiment. In our next study, we then set out to track the emergence of first impressions in the brain – this time from stimulus onset onwards in a passive listening paradigm (Lavan et al., 2024). We thus shifted the focus of the study away from the degree of exposure needed: Instead, we now homed in on when listeners seem to start forming first impressions when listening to a voice, without restricting their exposure to specific durations . We found partially convergent evidence for a similar staggered time course in this electroencephalogram (EEG) study: Using Representational Similarity Analysis, we found that representations of all person characteristics included in the study (sex, age, health, professionalism, educatedness, attractiveness, trustworthiness and dominance) could be decoded within less than 100ms of exposure of a voice (Lavan et al., 2024). This is remarkable, given that participants were at no point asked to actively evaluate the voices or people while listening to stimuli in the study. This therefore suggests that listeners routinely and perhaps spontaneously form impressions, even when not explicitly asked to do so. In line with models of voice processing and other EEG studies (Ambrus et al., 2019; Belin et al., 2011; Schirmer & Kotz, 2006), we found that these representations were earlier on (before ~400ms after stimulus onset) shaped by acoustic properties of the voice, to later become invariant to these acoustic properties. Intriguingly, when looking for independent representations of individual characteristics (after accounting for e.g. acoustic and other perceptual properties), we found that sex and age impressions can be decoded earlier than independent representations of dominance, professionalism and trustworthiness impressions.

Although the findings of the behavioural and EEG studies are not perfectly aligned, both studies therefore crucially show a staggered time course where some characteristics are perceived more quickly than others. The staggered time courses we observe may provide insights into how the multivariate first impressions that we form from voices are put together (Lavan & McGettigan, 2023): Could it, for example, be that we perceive a small number of person characteristics first, with characteristics that are perceived later being increasingly susceptible to top-down influences? That is, if we first form an impression of the sex and age of a person, could these initial impressions facilitate or inhibit which kinds of other characteristics we might ascribe to a person? If this were the case, the staggered time courses we observed could offer part of an explanation of how top-down influences of social stereotypes, over-generalisation effects, and halo effects (Forgas & Laham, 2016; Irvine & Lavan, 2025; Montepare & Dobish, 2003; Nisbett & Wilson, 1977) come to shape first impressions.

The section above highlights some of the recent developments in impression formation, which have provided intriguing glimpses into how listeners come to infer impressions from voices. However, much more work is needed before we can fully understand how multivariate first impressions are formed for voices (and faces). In such future work, we may want to assess how bottom-up and top-down processes are traded off against each other when forming impressions, how much stereotypes shape first impressions, whether (harmful) influences of stereotypes on impressions can be counteracted, and find out more about how and to what degree impressions influence and guide our daily behaviour and decisions.

Encountering a voice time after time: Identity learning and identity recognition from voices

Alongside hearing unfamiliar voices that we have never heard before, we will also encounter voices that we have heard many times before, such that the voices have become familiar to us. In contrast to unfamiliar voices, listeners can recognise familiar voices and can therefore identify the specific identity of a known person. If listeners therefore want to make sense of who a person is, when hearing a familiar voice, recognising that specific voice will achieve this goal as it links the heard (and recognised) voice to all the information about that person that we have stored in our memory.

Factors affecting the accuracy of voice identity recognition: Familiarity and individual differences

How well listeners can recognise other people from their voices is difficult to say in absolute terms: Recognition accuracy across different studies depends significantly on the specific task and voices investigated. It is therefore often more meaningful to express listeners’ ability to recognise voice identity in relative terms. Research comparing the accuracy of voice identity recognition vs face identity recognition often suggests that face recognition is more accurate and more robust to perceptual challenges than voice recognition (Barsics, 2014; Stevenage & Neil, 2014; Young et al., 2020). Furthermore, when looking at how successfully listeners extract other kinds of information from voices, we can see that listeners are experts at extracting speech information from voices, while voice identity recognition again appears to be comparatively less accurate and robust overall (Young et al., 2020). Thus, although people can recognise others from voices only, voice identity recognition is not one of the perceptual skills that humans excel at in most circumstances.

Having said that, there are situations and circumstances in voice identity perception is extremely robust and accurate: While listeners may struggle to recognise voices they do not know that well, we have found that listeners can recognise personally familiar voices (in this case a romantic partner of more than 6 months) with extremely high accuracy, even in very challenging listening tasks (Kanber et al., 2022). For example, listeners can recognise the voices of their romantic partners with almost perfect accuracy from conversational filler sounds (‘uhm’, ‘uhhh’, ‘ehm’) and recordings that had been altered through drastic acoustic manipulations. The same listeners were, however, only able to recognise voices they had been trained to recognise in the lab with low-to-moderate accuracy in the very same tasks. These studies therefore show that listeners can be extremely good at voice identity perception under some circumstances, although extensive exposure and engagement with a voice, that goes well beyond, for example, lab-based training, is necessary to achieve this.

Other work confirms that the degree of familiarity and type of exposure matter for how well people can recognise voices: In an earlier study (Lavan et al., 2016), we recorded a group of university lecturers either producing steady-state vowels or laughter. We then asked students who had either been taught by these lecturers and students who had not been taught by them and had never heard them elsewhere to complete a voice discrimination task. While being familiar with the voices led to overall better performance, the familiarity that the students had with the lecturers’ voices was not sufficient to yield the extremely robust ceiling performance we observe when testing listeners on the voices of their romantic partners (Kanber et al., 2022). Instead, listeners who were familiar with the lecturer’s voices still made many errors when making voice discrimination judgements when the task was difficult (e.g. when asked to make judgements for a pair of recordings including a vowel and a laugh, two vocalisations that are quite different from one another and thus harder to compare (Lavan, Burton, et al., 2019). This shows that even though students had extensive exposure to the lecturers’ voices and were thus quite familiar with them, the type of exposure was still not sufficient to complete the challenging aspects of the task with close to no errors. We might expect (but unfortunately neglected to test in our later study) that this difficult voice discrimination task would not be a challenge if we had sampled the voices of the participants’ romantic partners.

Some studies systematically vary the amount of exposure or training listeners get to become familiar with a voice, thus precisely controlling how familiar(ised) a person is with a voice. In these studies, evidence for how increasing familiarity affects identity perception is somewhat more mixed. Kanber (2022, Experiment 5) reports that participants who learned voices from more versus less training (learning from 80 recordings vs. 20 recordings) performed better at a voice identity recognition task after training. Effect sizes for this improvement were, however, relatively modest (Hu score of 0.54 for more training compared to 0.46 for less training). Conversely, in a study using a similar manipulation, Holmes et al. (2021) did not find similar increases in voice identity recognition accuracy after longer training, when contrasting voices trained on 88, 166 and 478 recordings, respectively.

Aside from factors such as the degree of familiarity, there are substantial individual differences in how well different people can recognise voices: At one extreme are people with phonagnosia (e.g. Garrido et al., 2009; Roswandowitz et al., 2017; Van Lancker et al., 1988). These individuals have selective problems with recognising familiar others from their voice, while still being able to discriminate between two unfamiliar voices, read emotions from voices or perceive speech information from voices. At the other extreme are so-called ‘super-recognisers’, whose voice recognition abilities, as measured by various tests of voice recognition, voice learning and voice discrimination, are several standard deviations above the population mean (Aglieri et al., 2017; Jenkins et al., 2021).

Identity learning: Establishing a representation to recognise and identify people from voices

There is a substantial body of research examining how listeners move from being unfamiliar with a voice to becoming familiar and thus able to recognise a voice. This literature has often focused on asking questions about how people form representations of specific voices when becoming familiar with them. Theoretical models have proposed a mechanism for voice learning and representation formation, where, following repeated exposures, listeners gradually lay down a mental representation of a specific voice identity (Maguinness et al., 2018). While after an initial exposure to a previously unfamiliar voice, listeners have no robust representation (or ‘stored deviance patterns’ in the terminology of the prototype models as invoked by Maguinness et al., 2018) of that voice, every repeated exposure adds to establishing a representation of an increasingly familiar voice.

We have shown that these representations of voices appear to be formed based on acoustic averages of a voice (Lavan, Knight, & McGettigan, 2019). Specifically, we trained participants to learn voices based on voice recordings that systematically varied in their acoustic properties. Participants were trained to recognise these identities based on stimulus distributions forming a ring-shaped distribution in a 2-dimensional acoustic voice space. Crucially, these distributions were missing their centres or acoustic averages. At test, listeners were better at recognising a trained voice when our test stimuli overlapped with the previously unheard average of each ring-shaped distribution, compared to test stimuli falling onto the trained ring-shaped distribution. Listeners were thus better at recognising trained voices from their acoustic averages, despite having never heard these specific averages before when learning the voices. We concluded that listeners form representations of familiar voices based on abstracting acoustic averages, which are then, in turn, used to recognise that familiar voice.

Furthermore, we have shown that voice identity learning and voice identity recognition are affected by how much variability listeners are exposed to when learning to recognise a voice. Building on the literature on non-native speech sound learning (Brekelmans et al., 2022; Lively et al., 1993; Logan et al., 1991) and face identity learning (Burton et al., 2016; Murphy et al., 2015; Ritchie & Burton, 2017), and beyond (Raviv et al., 2022 for a review), we asked whether high-variability training would lead to more accurate and more generalisable learning of voice identities than low-variability training (Lavan, Knight, Hazan, et al., 2019). High-variability training was defined as training where listeners were familiarised with voices based on stimuli produced in several different speaking styles (reading sentences, spontaneous conversational speech, speaking to be maximally intelligible, etc.), compared to low-variability training, which included stimuli from only one of these speaking styles. Across three experiments, we saw some evidence supporting our hypothesis, while also getting insights into situations where high-variability training may indeed be a disadvantage to listeners.

High-variability training only led to more accurate voice perception when the test stimuli were sampled from a previously unheard speaking style. That is, high-variability advantages only emerged when listeners had to generalise to something they had no direct experience with before. Being able to generalise to novel instances of what a person’s voice sounds like is particularly important to being successful at voice identity perception in our everyday lives. We constantly change the sound of our voice to adapt to different situations (Lavan, Burton, et al., 2019): Our voice changes depending on our mood (e.g. happy vs. sad), our conversation partners (e.g. a close friend vs. our boss), the speaking environment we are in (a quiet room vs. a noisy bar), and more. To successfully cope with this flexibility and variability in our voices, listeners need to link (Lavan, Burston, et al., 2019; Lavan, Burton, et al., 2019) a particular version of a voice they are hearing – sometimes a version of a person’s voice that they may not have heard before – to the representation of that voice that they have built up over time.

While high-variability exposure helped in situations where generalisation beyond what was already familiar was important, we also found that high-variability training is not always beneficial: When listeners were trained in the low-variability condition on the same speaking style that they were then tested on, low-variability advantages emerged, given the more targeted, relevant exposure that the low-variability training offered. Interestingly, we also observed in one experiment that if participants were given only a few stimuli per voice in the context of high-variability training, participants were altogether unable to learn some voices with above-chance accuracy during training. This suggests that, in the context of sparse but highly variable input, it can be too challenging for participants to integrate high-variability stimuli that all sound quite different from one another into a single coherent voice identity, making learning impossible. This observation thus shows that high-variability training can under specific circumstances indeed be detrimental.

The section above illustrates that, while voice identity perception is relatively less robust and accurate than face identity perception, listeners can still make sense of which specific, familiar person they are talking to via identity recognition. Studies have highlighted that becoming familiar with a person requires the listeners to form a (robust) representation of a familiar person’s voice, which can then be used to recognise the voice later on. What the content of these representations is and how different representations are organised in relation to one another is not yet well understood, such that more work is required to understand how these representations that support voice identity recognition are structured and built.

Independent literatures, co-dependent percepts? Integrating first impression formation and identity recognition

To date, there has been only limited cross-talk between work on voice impression formation and voice identity recognition and learning. However, if we assume that the main aim of person perception from voices is to make sense of who we are hearing, we can potentially look at impression formation and identity recognition within a single conceptual framework. We can position first impression formation and identity recognising on two ends of a familiarity continuum (Figure 1), which enables us to generate interesting and to date largely untested research questions, by moving away from the endpoints of this familiarity continuum: What happens to our first impressions when we hear a voice for more time? After how much exposure to a voice and with how much familiarity does identity-specific perception take over from impression formation? Can these two ways of making sense of voices support each other and interact for voices that are neither completely unfamiliar nor extremely familiar to us?

Figure 1.

Illustration of a continuum or gradient ranging from first impressions to robust recognition of voices.

Beyond first impressions: How do first impressions become second and, eventually, lasting impressions?

We have seen in previous sections of this paper that first impressions can be formed based on often minimal information, with participants relying on both top-down information (e.g. social stereotypes) and perceptual information (e.g. low pitch) to form these impressions. However, what happens to our first impressions when we get more information? First impressions are likely fleeting as they are mainly based on heuristics, such that we should update these first impressions as soon as new, useful information becomes available to us. Interestingly, there is not much work in the voice perception literature that has tracked how impressions evolve after an initial encounter. What listeners do when they hear a voice over a longer period of time or across a number of different encounters has been primarily studied in the context of voice identity learning.

Some recent research shows that first impressions can be updated from cross-modal information: Masi et al. (2022) show that first impressions based on unfamiliar voices change after listeners are shown a picture of an (unfamiliar) face afterwards (and vice versa). Intriguingly, Masi et al. (2022) report that voices have a greater impact on impressions than faces do. Similarly, we familiarised listeners with voices via a valenced training paradigm by giving them positively or negatively valenced information about the behaviours of the voices. Compared to the listeners providing first impressions, the listeners who had undergone the familiarisation rated the voices more positively or negatively, in line with their training experience (Lavan, Mileva, & McGettigan, 2021) This shows that the valenced training potentially shaped their impressions. Taken together, both studies provide some initial evidence that first impressions do get updated in light of any other information (here: a face and written context) that is or becomes available to listeners.

However, to our knowledge, there are to date no studies that track how first impressions from voices are updated when listeners are exposed to more voice-based information (e.g. a second voice recording that may convey slightly different [first] impressions. Furthermore, no research examines how impressions are updated dynamically, across more than the two time points sampled in the studies above. There are, however, different possibilities for how impressions may evolve over time. For example, it may be possible that impressions are initially relatively volatile, such that any new information can potentially have a large impact on how we evaluate a relatively unfamiliar person. Over time and after sampling more information, any new information may gradually have less of an impact on the ever more stable impression we have formed of this increasingly familiar person. Similarly, we do not yet know how the potentially volatile early impressions of a person may settle out and converge on a more stable impression: Would the impression, for example, converge on an average impression of our previous experiences or something else (see voice identity representations)? How would ‘outlier’ behaviour that is out of character for a person be treated when updating impressions that are more or less stable?

It is also an open question when and how voice identity recognition takes over the task of ‘making sense of a person’ from impression formation after a listener becomes ever more familiar with a person because being able to access specific, stored information about a person via recognising their identity provides a better, more reliable source of person-related information than the heuristics and general social stereotypes that are used to form first impressions in the absence of other information. Furthermore, when a point of familiarity is reached, where identity perception has taken over ‘making sense of a person’, what happens to impression formation for that voice? Are impressions still formed and updated when listening to familiar voices, but now with the purpose of making sense of ad-hoc changes in a person’s behaviour instead of trying to infer stable characteristics? Our study, looking at whether impressions from hearing multiple recordings of familiar voices are more stable than impressions from unfamiliar voices, may suggest that this is the case (Lavan, Mileva, & McGettigan, 2021): Despite being familiar with the voices tested, listeners’ impressions were as variable as the impressions formed by unfamiliar listeners.

Before voice identity recognition: Can first impressions support or shape voice identity learning?

While there is curiously little research on how first impressions evolve over continued exposure and increasing familiarity, we have seen that there is a substantial literature on voice identity learning. This literature, however, defines the starting point of voice identity learning as listeners being unfamiliar with a voice, such that they cannot recognise that voice. This is, of course, correct: Unfamiliar voices cannot be recognised. However, this view overlooks that unfamiliar voices do give rise to first (and second) impressions, which provide listeners with person-related information. After matching the to-be-learned voices for demographics, accents, and sometimes broad acoustic properties, voices (within an experimental condition) are then treated as being interchangeable by researchers for the purpose of identity learning. This approach, however, overlooks that the first impressions formed from these voices may not be well matched at all across the voices studied.

It might be reasonable to assume that the first impressions different voices evoke could affect and shape voice identity learning. There may already be indirect evidence for such effects: Research has already shown that increasing voice distinctiveness affects voice memory (Bülthoff & Newell, 2015; Kreiman & Papcun, 1991; Stevenage et al., 2018). ‘Distinctiveness’ is often defined as the distance from an average voice in a voice space and is usually measured through perceptual ratings (e.g. ‘How distinctive is this voice?’). Many studies do not specify which perceptual features contribute to perceptions of distinctiveness. Those that do often link distinctiveness to psycho-acoustic properties of the voices (e.g. specific voice qualities, particular ways of speaking). Crucially, however, another study shows that differences in perceived distinctiveness in unfamiliar voices can also affect impressions of attractiveness and likability (Zäske et al., 2020). From this, it follows that if distinctiveness influences voice identity learning and if distinctiveness is related to first impressions, then first impressions should also impact voice identity learning. It is unclear whether distinctiveness (as a perceptual characteristic of the sound of a voice) or first impressions (as inferred characteristics of a person) drive and underpin these potential influences on voice identity learning, or whether both types of perceptual features contribute in some way. With there to date being only indirect empirical evidence for effects of first impressions on later voice learning, new data is needed to start to test this question, such that any findings can shape our theories of how voice learning is achieved, and potentially scaffolded by other voice-related perceptual features.

Recognising voices that are not so familiar – via (accurate) person perception and impressions

Going a step further, we can speculate whether impressions and/or the perception of person characteristics may also scaffold voice recognition from voices that listeners are only moderately familiar with.

Models of voice perception have posited that voice processing follows processing steps, starting with a low-level acoustic analysis of the heard sound, followed by a stage called the ‘voice structural analysis’ in which voice sounds are identified and encoded, after which familiar voice identities are eventually recognised (Belin et al., 2004, 2011; Campanella & Belin, 2007). Other models have claimed that familiar voices are recognised more or less directly from the acoustic input based on ‘familiar voice patterns’ (Kreiman & Sidtis, 2011; Sidtis & Kreiman, 2012). We recently proposed that while these more or less direct pathways from acoustic input to identity recognition can exist, perhaps in particular for very familiar voices, it is also possible that identity recognition can be achieved via a more indirect route for some less familiar (and thus often less easily recognised) voices (Lavan & McGettigan, 2023). Voice identity recognition in these cases may be supported by listeners first perceiving some person characteristics from the voice, without initially recognising the identity. Over exposure time, and as a result of collecting more information about the person (from the sound of their voice or other sources) for an encounter, voice identity recognition can then be achieved later on. Arguably, given how rapid and often accurate, for example, sex perception from cisgendered people can be (Lavan, 2023b; Owren et al., 2007), it is also conceivable that sex perception may precede identity perception, which is likely somewhat slower. If sex information (and beyond) were routinely available to listeners when making identity judgements, we could speculate that, in fact, any available person characteristics might help listeners narrow down and zone in on relevant voices during familiar voice identity perception.

Conclusion

This paper highlights how studying identity perception, person perception, and impression formation separately from one another has potentially restricted the kinds of research questions we have been asking to examine how listeners make sense of who they are hearing. When putting the perception of these different kinds of person information into a single framework, by trying to explain their perception through a single process of recognition (Lavan & McGettigan, 2023) or by locating them along a single continuum of exposure/familiarity, we can find that identity perception and multivariate person perception indeed havemuch in common. Bringing them together and examining how they interact, support each other, and where one may limit the other, thus offers many fruitful avenues for future research.