The Time Course of Person Perception From Voices: A Behavioral Study

Abstract

Listeners spontaneously form impressions of a person from their voice: Is someone old or young? Trustworthy or untrustworthy? Some studies suggest that these impressions emerge rapidly (e.g., < 400 ms for traits), but it is unclear just how rapidly different impressions can emerge and whether the time courses differ across characteristics. I presented 618 adult listeners with voice recordings ranging from 25 ms to 800 ms in duration and asked them to rate physical (age, sex, health), trait (trustworthiness, dominance, attractiveness), and social (educatedness, poshness, professionalism) characteristics. I then used interrater agreement as an index for impression formation. Impressions of physical characteristics and dominance emerged fastest, showing high agreement after only 25 ms of exposure. In contrast, agreement for trait and social characteristics was initially low to moderate and gradually increased. Such a staggered time course suggests that there could be a temporo-perceptual hierarchy for person perception in which faster impressions could influence later ones.

Both personal and shared taste shape impressions from voices and faces

“…Participants rated 100 short recordings of voices and images of faces for one of 8 person characteristics (gender, age, health, attractiveness, dominance, competence, trustworthiness, likeability). These person characteristics were chosen to sample a range of primarily physical (e.g., gender, age) and being primarily social (e.g., likability, trustworthiness) characteristics that have been shown to be important to voice (and also face) perception (Lavan, 2023;McAleer et al, 2014;Oosterhof & Todorov, 2008;Sutherland et al, 2013;Lin et al, 2021). We then used a variance portioning approach to quantify the role of shared taste (i.e., aspects of impressions formation that different listeners generally agree on, such as lower pitched voices sounding more masculine) relative to personal taste (i.e., aspects of impressions formation that are idiosyncratic to a listener, such as one person preferring husky voices, while another prefers clearer voices).…”

“…For gender, 1 = very feminine and 9 = very masculine; for age 1 = "sounds/looks like a young adult" and 9 = "sounds/looks like an old adult"). These particular rating scales were chosen to cover a range of physical and social traits that are frequently measured in the existing literature of voice and face perception (Hehman et al, 2017;Lavan, 2023;McAleer et al, 2015;Oosterhof & Todorov, 2008;Sutherland, Burton et al, 2020).…”

“…Thus, 30-32 participants provided ratings for each person characteristic. The sample sizes were modelled on previous studies of trait perception (e.g., Hehman et al, 2017;Hönekopp, 2007;Lavan, 2023;Mileva & Lavan, 2023;McAleer et al, 2014). All participants were recruited via Prolific.co and were native speakers of German (matching the language spoken in the voice stimuli, see Materials), aged 18-65, no self-reported hearing impairments (for voices only), and had normal or corrected-to-normal vision (for faces only).…”

“…Research on trait perception from voices usually stresses that impressions are shared across listeners (e.g., McAleer et al, 2014;Mileva & Lavan, 2023, Lavan, 2023Rezlescu et al, 2015;Mahrholz et al, 2018): That is, listeners are suggested to agree with one another on whether a person sounds, for example, pleasant, friendly, feminine or older (or not). This focus on the shared nature of impressions has in turn driven theory development, such that theories usually emphasise that the apparent agreement between perceivers could be evidence speaking to listeners judgements being based on either common (social) stereotypes (e.g., Aronovitch, 1976;Schirmer et al, 2020;Zuckerman & Driver, 1990) or posit that perception and production have been shaped by evolutionary processes (e.g., Puts et al, 2006;Pisanski & Bryant, 2019;Hughes & Rhodes, 2010).…”

Reconstructing Voice Identity from Noninvasive Auditory Cortex Recordings

“…Here, we addressed this challenge by training a 'Variational autoencoder' (VAE; Kingma et Welling, 2014) DNN to reconstruct voice spectrograms from 182,000 250-ms voice samples from 405 different speaker identities in 8 different languages from the CommonVoice database (Ardila et al, 2020). Brief (250 ms) samples were used to emphasize speaker identity-related information in voice, already available after a few hundred milliseconds (Schweinberger et al, 1997;Lavan, 2023), over linguistic information unfolding over longer periods (word, >350 ms;Mcallister et al, 1994). While a quarter of a second is admittedly short compared to standards of, e.g., computational speaker identification that typically uses 2-3 s samples, this short duration is sufficient to allow near-perfect gender classification and performance levels well above chance for speaker discrimination (Fig.…”