Individual differences in human voice pitch are preserved from speech to screams, roars and pain cries

Abstract

Fundamental frequency (F0, perceived as voice pitch) predicts sex and age, hormonal status, mating success and a range of social traits, and thus functions as an important biosocial marker in modal speech. Yet, the role of F0 in human nonverbal vocalizations remains unclear, and given considerable variability in F0 across call types, it is not known whether F0 cues to vocalizer attributes are shared across speech and nonverbal vocalizations. Here, using a corpus of vocal sounds from 51 men and women, we examined whether individual differences in F0 are retained across neutral speech, valenced speech and nonverbal vocalizations (screams, roars and pain cries). Acoustic analyses revealed substantial variability in F0 across vocal types, with mean F0 increasing as much as 10-fold in screams compared to speech in the same individual. Despite these extreme pitch differences, sexual dimorphism was preserved within call types and, critically, inter-individual differences in F0 correlated across vocal types (r = 0.36–0.80) with stronger relationships between vocal types of the same valence (e.g. 38% of the variance in roar F0 was predicted by aggressive speech F0). Our results indicate that biologically and socially relevant indexical cues in the human voice are preserved in simulated valenced speech and vocalizations, including vocalizations characterized by extreme F0 modulation, suggesting that voice pitch may function as a reliable individual and biosocial marker across disparate communication contexts.

1. Introduction

In addition to being the carrier of language, the human voice has been shaped by selection to communicate biologically relevant traits of the vocalizer ([1] for review). Indexical cues to a vocalizer's identity, sex and age are readily transmitted by the non-linguistic properties of modal speech, particularly fundamental frequency (F0), perceived as voice pitch ([2] for review). Fundamental frequency in modal speech corresponds to the rate at which the vocal folds vibrate, determined by their effective mass, length and tension. In addition to fluctuating dynamically during speech production within a single vocalizer ([3,4] for reviews), F0 varies widely across individuals, both between the sexes (men's voices are lower-pitched than women's) and within adults of the same sex, due to a combination of endocrinological, physiological, anatomical and social factors [5].

Voice pitch plays an important role in human social interactions. In addition to reliably indicating sex and age, individual differences in the mean pitch of speech utterances facilitate speaker recognition, predict mating success and social status, and influence listeners' biosocial judgements of vocalizers ([2,6,7] for reviews). However, past work in this area has focused almost exclusively on the role of pitch in speech, largely ignoring the variable nature of F0 across a much wider range of human vocal sounds, and thus overlooking the complexity of the human vocal apparatus and its evolved functions.

Humans can readily modulate their voice F0 to express a range of emotions and motivations during speech production [3,4], and studies conducted in real-life contexts have shown that men and women alter their speech F0 depending on social context or to whom they are speaking (see e.g. [8,9]). Yet beyond verbal communication, humans also frequently produce non-linguistic vocalizations ranging from laughter to screams that also contain salient F0 cues of potential biological and social relevance [10]. Such vocalizations almost certainly emerged in the human vocal repertoire before words [3], and can exploit a broader, less constrained acoustic space than modal speech. Indeed, intelligible speech production necessitates a relatively steady F0 ensured by an ability to maintain constant subglottal pressure [2], as well as a dense harmonic structure (thus relatively low F0) to effectively encode phonetic information, including discriminable formant frequencies [11]. By contrast, human nonverbal vocalizations, which can resemble those of non-human mammals (e.g. [12]), are often characterized by extreme fluctuations in voice pitch and, like those of other mammals, their acoustic structure typically maps onto their purported biological or social function [13]. This form-function mapping across vocal types appears to function to exploit listeners’ perceptual sensitivities and biases. For example, distress vocalizations (cries, screams) are typically several orders of magnitude louder and higher-pitched than modal speech, attracting attention and eliciting arousal [14]. Such distress vocalizations also exhibit different spectral profiles than those of agonistic or aggressive vocalizations (roars, grunts), which are typically characterized by a relatively low F0 and a high proportion of nonlinear phenomena (e.g. [15,16]).

Given this considerable variability, an important question is whether acoustic cues to vocalizer attributes are shared across call types (e.g. as observed in red deer [17]) or are call-specific (e.g. as observed in zebra finches [18]). Indeed, in humans, no previous study has examined whether individual differences in F0 are preserved across the broad range of human vocal sounds, from modal and emotionally valenced speech to nonverbal vocalizations. While it is possible that idiosyncratic differences in F0 between different vocalizers might be overridden by the more extreme F0 modulations that characterize agonistic and distress vocalizations, existing data on the F0 profiles of human grunts [15], roars [16], laughs and cries [19] indicate that such vocalizations retain a degree of sexual dimorphism, wherein men produce relatively lower-pitched vocalizations than do women. There is also preliminary evidence that within each sex, F0 in modal speech correlates with F0 in sung speech [20], and that cues to individual identity are retained in valenced human speech [21], laughter [22], cries [23], and in the screams of both humans ([24], cf. [25]) and non-human primates [26] (with the caveat that speaker recognition is substantially reduced from these vocalizations compared to modal speech among human listeners [21,22,25]). Finally, longitudinal studies have recently revealed that individual differences in F0 remain relatively stable across the lifespan, from infancy to childhood and throughout adulthood [27,28]. Taken together, this body of literature suggests that F0 may represent a reliable and surprisingly stable individual marker despite its extreme dynamicity at the within-vocalizer level (e.g. across vocal types). From an evolutionary and ethological perspective, such F0 stability could function to provide honest information about a vocalizer's identity and biosocial profile (e.g. dominance, masculinity) regardless of the mode of vocal production, or social context.

Here, we investigate the stability of individual differences in F0 across speech and nonverbal vocalizations produced by the same men and women in neutral, aggressive, fearful and pain contexts, by contrasting within-individual variation in F0 between and across these diverse call types.

2. Material and methods

2.1. Participants

We recorded the voices of 51 drama or acting students from the Royal Central School of Speech and Drama (London, UK) and the University of Sussex (Brighton, UK). The sample included 28 men (mean age 23.6, range 18–71) and 23 women (mean age 22.8, range 18–66) who provided informed consent and received monetary compensation. No participant reported conditions that could affect their voice (e.g. cold, sore throat, chronic smoking).

2.2. Voice recording

Participants were audio recorded in a quiet room using a Zoom H4n microphone placed at a distance of 150 cm to prevent audio clipping. Each vocalizer produced eight simulated ‘vocal types’: neutral speech, aggressive speech, fearful speech, an aggressive vocalization (i.e. ‘roar’), a fearful vocalization (i.e. scream), and pain vocalizations corresponding to mild, moderate and intense levels of pain (figure 1). Voice recordings were saved as WAV files at 44.1 kHz sampling frequency and 16-bit resolution, and later transferred to a laptop for acoustic analysis.

Figure 1.

Waveforms and spectrograms representing each of eight vocal types produced by a single representative individual (female, aged 22), demonstrating the high degree of intra-individual variability in F0 across vocal types.

To obtain neutral modal speech, participants produced the first sentence of the Rainbow Passage [29] in a modal voice. They were then instructed to imagine themselves in various scenarios and to produce speech sentences and/or vocalizations to express their motivation and emotion in each given context. A description of the context or speech sentence was dictated by the experimenter and also displayed on a computer screen. In the aggression and fear contexts, participants imagined themselves on a battlefield, either attacking or being attacked, and produced the sentences, ‘That's enough, I'm coming for you!’ and, ‘Please, show mercy, don't hurt me!’, respectively. They were further instructed to produce a nonverbal vocalization to express that same motivation. In the pain context, participants produced vocalizations in three imagined scenarios of increasing pain intensity: mild (e.g. stubbing their toe), moderate (e.g. breaking their arm) and intense (e.g. childbirth). Full context descriptions are given in Raine et al. [16,30].

2.3. Acoustic analysis

Acoustic editing and analysis were performed in PRAAT 5.3.62 [31]. Recordings were segregated by vocal type, resulting in 408 audio clips, and edited manually to remove silences or acute background noises. Fundamental frequency (F0) was measured using a custom script with a search range of 60–2000 Hz, 0.05 s window length and 0.01 time step. Extracted F0 contours were systematically inspected and verified, and any measurement errors (e.g. arising from octave jumps, sub-harmonics or deterministic chaos) were de-selected or corrected. This established method has been applied successfully in a number of studies to measure F0 both in human speech and nonverbal vocalizations characterized by extreme F0 values (e.g. babies cries [32], tennis grunts [15]). Measured F0 was converted from hertz (Hz) to equivalent rectangular bandwidths (ERBs), a quasi-logarithmic scale that controls for any discrepancy between measured F0 and perceived voice pitch.

2.4. Statistical analysis

A linear mixed model (LMM) fit by restricted maximum-likelihood estimation was first used to test for differences in F0 (ERB) across vocal types and between sexes. Vocal type and sex of vocalizer were entered into the omnibus model as fixed variables, and vocalizer ID was entered as a random variable with random intercept. This was followed by separate LMMs for each vocal type to more closely examine sex differences. Significant effects were examined using pairwise tests with Šidák correction for multiple comparisons.

Correlation matrices (Spearman's rho, r_s, one-tailed) were then conducted to test for positive relationships in F0 between all vocal types, separately for each sex and each F0 scale (ERB, Hz). Shapiro–Wilk tests indicated that F0 was not normally distributed in nonverbal vocalizations (see electronic supplementary material, table S1), hence non-parametric tests were used for regression analyses. Bootstrapping was used to compute 95% confidence intervals for each bivariate correlation. The Benjamini–Hochberg procedure ([33], where m = 28, q = 0.1) was employed to control for the inflated false discovery rate owing to multiple comparisons. Datasets and statistical scripts are provided as electronic supplementary material.

3. Results

3.1. Sexual dimorphism in F0 across vocal types

Figure 2 illustrates distributions in mean F0 across the eight vocal types for each sex. For illustrative purposes F0 values are plotted in Hertz along the y-axis and means are given in both Hertz and ERBs above each violin plot.

Figure 2.

Violin plots representing the full distribution in mean F0 for each sex and each vocal type. Mean F0 values are given in both Hertz and ERBs above the violin plot for each sex and each vocal type. Significant sex differences were observed for each vocal type (***p < 0.001 following Šidák correction). Plots were produced in R ggplot 2 package.

The omnibus LMM was significant (intercept: F_1,49 = 4949.6, p < 0.001) and showed significant effects of vocalizer sex (F_1,49 = 123.1, p < 0.001), vocal type (F_7,343 = 121.4, p < 0.001) and a significant interaction between sex and vocal type (F_7,343 = 4.1, p < 0.001) on mean F0.

Separate LMMs conducted for each vocal type confirmed that men's mean F0 was significantly lower than women's for each vocal type, including modal speech, all valenced speech and nonverbal vocalizations (all F_1,49 > 23.7, all p < 0.001; see electronic supplementary material, table S2 for pairwise comparisons, where all p < 0.001 following Šidák correction). Sexual dimorphism was particularly pronounced for nonverbal vocalizations, wherein the F0s of women's fear screams and aggressive roars were on average 4.5 ERBs (433 Hz) and 3 ERBs (241 Hz) higher than men's, respectively. The F0s of women's pain cries ranged from an average of 2.6–3.2 ERBs (194–297 Hz) higher than men's (figure 2; electronic supplementary material, table S2).

3.2. Stability of individual differences in F0 across vocal types

Within each sex, significant moderate to strong positive relationships between F0s (r = 0.36–0.80) were also observed across various vocal types (table 1; see electronic supplementary material, table S4 for comparable results in Hz). For significant correlations, the mean F0 of a given vocal type explained between 13 and 64% of the variance in the mean F0 of another vocal type, on average, within the same sample of same-sex vocalizers (table 1).

Table 1.

Correlation matrix of F0 (ERB), pairwise comparisons across all vocal types. Results for men's voices are given in the top right portion of the table (significant relationships are highlighted in blue), and results for women's voices are given in the bottom left portion of the table (significant relationships are highlighted in beige). See electronic supplementary material, table S3 for exact p-values and electronic supplementary material, table S4 for analogous comparisons on Hz scale. See electronic supplementary material, figure S1 for corresponding scatterplot matrices.

	statistic	modal speech	aggress speech	fear speech	aggress roar	fear scream	pain mild	pain mod	pain intense
modal speech	rs		−0.06	0.36*	0.16	0.15	0.80**	0.31	0.39*
	95% CI		−0.41, 0.30	−0.07, 0.68	−0.24, 0.50	−0.29, 0.53	0.63, 0.89	−0.12, 0.65	0.01, 0.69
aggress speech	rs	0.09		0.55**	0.43*	0.26	0.11	0.08	0.15
	95% CI	−0.33, 0.47		0.15, 0.77	0.04, 0.75	−0.13, 0.53	−0.30, 0.48	−0.28, 0.45	−0.22, 0.48
fear speech	rs	0.04	0.46*		0.41*	0.54**	0.50**	0.12	0.26
	95% CI	−0.47, 0.51	0.09, 0.75		0.03, 0.68	0.16, 0.77	0.10, 0.75	−0.23, 0.45	−0.12, 0.58
aggress roar	rs	0.23	0.62**	0.32		0.26	0.08	0.22	0.53**
	95% CI	−0.19, 0.58	0.30, 0.83	−0.06, 0.64		−0.14, 0.58	−0.29, 0.45	−0.18, 0.56	0.23, 0.74
fear scream	rs	−0.08	0.30	0.45*	0.30		0.26	0.28	0.25
	95% CI	−0.51, 0.38	−0.14, 0.71	0.10, 0.71	−0.13, 0.63		−0.18, 0.61	−0.14, 0.65	−0.13, 0.57
pain mild	rs	0.15	−0.17	−0.13	−0.01	0.55**		0.39*	0.28
	95% CI	−0.28, 0.48	−0.59, 0.31	−0.56, 0.29	−0.44, 0.47	0.20, 0.80		−0.03, 0.67	−0.12, 0.62
pain mod	rs	−0.07	−0.12	0.07	−0.17	0.25	0.43*		0.30
	95% CI	−0.49, 0.36	−0.51, 0.30	−0.40, 0.49	−0.60, 0.30	−0.13, 0.54	−0.01, 0.74		−0.11, 0.67
pain intense	rs	−0.06	0.18	0.12	0.03	0.64**	0.44*	0.38×
	95% CI	−0.45, 0.35	−0.25, 0.66	−0.36, 0.57	−0.42, 0.51	0.34, 0.81	−0.03, 0.80	−0.05, 0.70

Notes: Spearman's rho (r_s) correlation coefficients are followed by lower and upper 95% confidence intervals (bootstrapping, 1000 samples). Highlighted cells show significant relationships where **p < 0.01, *p < 0.05. A single relationship (marked ‘×’) did not pass Benjamini–Hochberg correction (unadjusted p = 0.04).

In both sexes, F0 measured from aggressive speech correlated strongly with F0 measured from fearful speech. Moreover, both of these valenced speech types predicted the mean F0 of corresponding nonverbal vocalizations, where the relationships were largely valence-specific: hence, aggressive speech F0 predicted roar F0, whereas fearful speech F0 predicted scream F0. Following the same valence-specificity, roar F0 did not significantly predict scream F0 in either sex (though these weak relationships approached significance, r = 0.26 and 0.30, p < 0.10). Within pain vocalizations, F0 correlated across at least two pain intensity levels in either sex, with significant relationships between mild and moderate or intense pain (table 1; electronic supplementary material, table S3).

While these results were generally consistent between the sexes, some sex differences emerged. The F0s of pain vocalizations correlated with roar F0 in men only, and with scream F0 in women only. Interestingly, F0 measured from modal speech was a relatively poor predictor of F0 in emotionally valenced speech or nonverbal vocalizations, particularly among women, with the exception that modal speech predicted men's fearful speech and pain vocalizations.

4. Discussion

We show that individual differences in human fundamental frequency (F0, perceived as voice pitch), previously observed in neutral speech and known to function as indexical and social signals in human conversation ([2,6,7] for reviews), are also present in valenced speech and in simulated non-linguistic vocal sounds, including those characterized by extreme F0 values (e.g. screams reaching nearly 2000 Hz). We further show that, despite a high degree of variability in F0 across eight different vocal types, between-individual differences in F0 are preserved across speech and vocalizations in a largely valence-specific manner. For instance, the mean F0 of men and women's aggressive speech sentences reliably predicted the mean F0 of their roars, with an analogous relationship between the F0s of fearful speech and screams. Individual differences in F0 were also preserved across pain vocalizations representing varying levels of pain intensity. All vocal types also retained sexual dimorphism and thus were significantly lower-pitched among men than women, with the most salient sex differences observed for screams, roars and pain cries.

Screams showed the most extreme F0 values overall (figure 1) and in turn the greatest degree of variability among vocalizers both between and within sexes. For example, the mean F0s of women's screams ranged from 484 to 1981 Hz, representing a difference of almost 1500 Hz (11 ERBs) between the highest- and lowest-pitched women. The fearful speech was by contrast much more constrained, with mean F0s ranging between 307 and 570 Hz in women (a maximum difference of 263 Hz or 3.7 ERBs between women). It is thus quite remarkable that individual differences in the F0s of valenced speech predicted those of such extreme vocalizations. By contrast, modal speech showed the lowest degree of F0 variability, with a maximum difference of only 112 Hz or 2.3 ERBs between women (F0 range 148–260 Hz) and 72 Hz or 1.8 ERBs between men (F0 range 81–153 Hz). Moreover, the F0 of modal speech was a poor predictor of the F0 of both valenced speech and nonverbal vocalizations, particularly among women for whom modal speech did not predict the F0 of any other vocal type. These results suggest that, in addition to valance-specificity, individual differences in F0 are also more strongly preserved within the broader category of emotional (compared to neutral) sounds.

These results complement two longitudinal studies that recently showed that the F0 of babies' cries predicts the F0 of their speech later in childhood [28], and that individual differences in F0 stabilize after puberty [27], suggesting that voice pitch is determined early in life and remains fairly stable throughout adulthood [27,28]. Those findings, together with the results of the present study and other recent work showing that individuals can generally be recognized when producing emotional speech [21], laughter [22], cries [23] and screams ([24], cf. [25]), suggest that the human voice functions as a stable individual marker despite its extreme dynamicity. With regards to voice F0, this may be due to anatomical constraints (probably linked to individual differences in vocal fold length) that maintain the relative inter-individual differences in F0 across the typical ranges for each vocal type. While few studies have investigated the stability of F0 differences across the vocal repertoires of other species, there is evidence that cues to individual identity may also be preserved across call types in non-human mammals, including red deer [17], rhesus macaques [34] and cows [35].

The results of the present study warrant replication on a larger, more representative sample, as the current study is unpowered with the small sample size of 28 men and 23 women. As a consequence, our analyses could fail to detect statistical significance in weak and even moderate correlations (r = 0.10–0.30). This is further illustrated by the broad confidence intervals observed for most correlations (table 1). Replication studies should also include individuals without acting experience, as the potential influence of vocal training on the spectral and perceptual dimensions of speech and nonverbal vocalizations remains unclear (see e.g. [36,37]). Finally, given the influence of social and cultural factors on voice production and vocal expressions of emotion, additional research is also needed to examine the generalizability of these results to other cultures.

Extending the present line of inquiry, follow-up studies may include a broader range of speech and vocalizations (e.g. positively valenced), and could test whether modal speech F0, which proved to be a poor predictor in the present study, correlates with the F0 of less extreme and comparatively ‘neutral’ vocalizations such as yawns. As researchers have observed some differences between simulated and spontaneous vocal sounds (e.g. identity-related information is more reliably encoded in genuine laughter [19,22]), future studies may also compare F0 relationships in volitional versus spontaneous vocal types, including naturally occurring vocal displays produced in real-life social contexts.

The key implication of stable inter-individual differences in voice pitch across vocal types is that honest biosocial cues about the speaker are likely to be communicated regardless of the mode of vocal production. Thus, our results also suggest that individual differences in F0 may be preserved across social (and emotional) contexts, though this hypothesis should be explicitly tested. This work also contributes to a small but growing body of research examining form and function in human nonverbal vocalizations, which, despite being understudied, constitute a substantial portion of the human vocal repertoire and promise to offer new insight into the evolutionary origins of vocal communication and the evolution of speech [3,10,12].

Ethics

The study was reviewed and approved by the Sciences and Technology Cross-Schools Research Ethics Committee (C-REC) of the University of Sussex (ER/JR307/4).

Data accessibility

The datasets and tables supporting this article are available as electronic supplementary material.

Authors' contributions

D.R., J.R. and K.P. designed the investigation. J.R. collected the data. J.R. and K.P. performed acoustic analyses. D.R. and K.P. performed statistical analyses. K.P. wrote the original manuscript and created figures and tables. K.P., J.R. and D.R. edited, revised and approved the final manuscript, and agree to be accountable for the work.

Competing interests

We declare we have no competing interests.

Funding

This work was partly supported by the University of Sussex through a scholarship to J.R. Author K.P. was supported by the University of Lyon IDEXLYON project as part of the ‘Programme Investissements d'Avenir’ (ANR-16-IDEX-0005) to D.R., and by the European Union's Horizon 2020 research and innovation programme under the Marie Skłodowska-Curie grant agreement no. 655859.