Human screams occupy a privileged niche in the communication soundscape

Abstract

Screaming is arguably one of the most relevant communication signals for survival in humans. Despite their practical relevance and their theoretical significance as innate [1] and virtually universal [2, 3] vocalizations, what makes screams a unique signal and how they are processed is not known. Here, we use acoustic analyses, psychophysical experiments, and neuroimaging to isolate those features that confer to screams their alarming nature, and we track their processing in the human brain. Using the modulation power spectrum (MPS [4, 5]), a recently developed, neurally informed characterization of sounds, we demonstrate that human screams cluster within restricted portion of the acoustic space (between ∼30 and 150 Hz modulation rates) that corresponds to a well-known perceptual attribute, roughness. In contrast to the received view that roughness is irrelevant for communication [6], our data reveal that the acoustic space occupied by the rough vocal regime is segregated from other signals, including speech, a pre-requisite to avoid false alarms in normal vocal communication. We show that roughness is present in natural alarm signals as well as in artificial alarms and that the presence of roughness in sounds boosts their detection in various tasks. Using fMRI, we show that acoustic roughness engages subcortical structures critical to rapidly appraise danger. Altogether, these data demonstrate that screams occupy a privileged acoustic niche that, being separated from other communication signals, ensures their biological and ultimately social efficiency.

Figure 1

The modulation power spectrum (MPS): examples and ecological relevance. (A) Representations of a 1000 Hz tone amplitude modulated at 25 Hz. Top, waveform. Middle, spectrogram. Bottom, MPS: power modulations in the spectral (y-axis) and temporal (x-axis) domains. 25 Hz modulation highlighted. (B) As in A. for a spoken sentence. (C) Modulations in human vocal communication. Perceptual attributes occupy distinct areas of the MPS and encode distinct categories of information. Modulations corresponding to pitch (blue) carry gender/size information [6, 11]. Temporal modulations below 20 Hz (green) encode linguistic meaning [13, 15]. Orange rectangles delimit ‘roughness’ [16, 17]. This unpleasant attribute has not yet been linked to ecologically relevant functions. We hypothesize that this part of the MPS space might be dedicated to alarm signals.

Figure 2

Acoustic characterization of screamed vocalizations. (A) Example spectrograms of the 4 utterance types, produced by one participant: screamed vocalizations, vowel [a] (top left); sentence (top right); neutral vocalizations, vowel [a] (bottom left); spoken sentence (bottom right). (B) Average MPS across participants (n=19) for each type. For the factorial analysis, the sentence factor (vertical dashed line) determines whether the utterance contains sentential information or the vowel [a]; the scream factor (horizontal dashed line) determines whether the utterance was screamed or neutral. (C) Main effect of scream. (D) Main effect of sentence. In B and C, contours delimit statistical thresholds of P < 0.001 (Bonferroni corrected).

Figure 3

Roughness modulations: natural and artificial sounds and behavior. (A) MPS roughness across categories. Left: screams, neutral speech, and musical (a capella) vocalizations. Center: artificial alarms versus musical instruments. Right: dissonant versus consonant sounds. (B) Perceived fear induced by natural and acoustically altered vocalizations. Left: Averaged rating (on a 1–5 negative scale) across participants, as a function of vocalization type: scream, filtered scream, neutral vocalization [a] and amplitude modulated (AM) neutral vocalization. Middle: Negative subjective ratings increase with MPS values in roughness range (red shading: 95% regression confidence interval). Right: average reaction times decrease with increasing roughness. (C) Spatial localization of screams, neutral vocalizations, and artificial screams. Left: localization accuracy. Center: speed. Right: efficiency. *** P < 0.001; ** P < 0.01; * P < 0.05; Error bars indicate SEM.

Figure 4

fMRI measurement of roughness and screams. (A) Main effect of unpleasantness across all sound categories. Unpleasant (rough) sounds induce larger responses bilaterally in the amygdala (left) and the primary auditory cortex (right). Contrasts are rendered at P < 0.005 threshold for display; see Table S4 for a summary of activations and associated anatomical coordinates. (B) Reverse-correlation analysis between single-trial beta values and MPS profiles of the corresponding sounds. The amygdala – but not primary auditory cortex – is maximally sensitive to the restricted spectro-temporal window corresponding to roughness. Contours delimit statistical thresholds of P < 0.05, cluster-corrected for multiple comparisons.