Terrifying film music mimics alarming acoustic feature of human screams

Music used to underscore frightening scenes in movies is often described as sounding “scream-like.” A well-known example is the music accompanying the infamous shower murder scene in Alfred Hitchcock's film Psycho (1960) with “screeching, upward glissandi” from the violins [Brown (1982), p. 46]. Although “scream-like” is a common descriptor, the question remains: do these scary film soundtrack excerpts actually sound like and are perceived similarly to human screams?

Music has a long history of relying on vocal behaviors to describe musical ones. For example, a branch of music theory (topic theory) has a categorical label (pianto) to describe music that mimics the sound of human weeping or sighs (Mirka, 2014). Recently, some music cognition researchers have begun to empirically investigate such instances of mimicry (Huron and Trevor, 2017; Trevor and Huron, 2019). These investigations are part of a branch of music and emotion research that theorizes that music might sometimes communicate emotion by mimicking human ethological vocal signals (Juslin and Laukka, 2003; Blumstein et al., 2012; Bryant, 2013; Huron, 2015; Warrenburg, 2019). Ethological signals are behaviors intended to communicate with a fellow member of one's species and cause them to react in a desired manner (Ehret, 2006; Lorenz, 1939). In humans, ethological signals can be smiling, crying, screaming, etc. (Huron, 2015; Ohala, 1996). Inspired by this branch of research and by the frequent comparison of scary music to human screams, the motivating question for the current study is: Do scream-like musical passages in scary film music actually mimic the sound of human screams to scare viewers?

What acoustic features characterize the sound of a human scream? Typically, human screams are loud, utilize a wide range of frequencies, are higher in pitch than one's average vocal range, and have a high amount of roughness (Arnal et al., 2015; Schwartz et al., 2019). Roughness is a basic auditory phenomenon that is characterized by a coarse, grating, or harsh subjective experience (Terhardt, 1974). Roughness has been defined in various ways, and a number of models and operationalizations have been formulated (Vassilakis and Kendall, 2010). Most models focus on amplitude modulation, especially in the broad region between 15 and 200 Hz. For the purposes of this study, we employ the modulation power spectrum (MPS) method and parameters used by Arnal et al. (2015). The MPS is a two-dimensional Fourier transformation of a soundwave that quantifies both temporal and spectral power modulations (Elliott and Theunissen, 2009). Previous research indicates that human screams feature higher MPS values than non-alarming vocalizations in the 30 to 150 Hz range of the temporal modulation rate dimension of the MPS (Arnal et al., 2015) [see Fig. 1(A)]. Vocal and artificial sounds exhibiting high temporal modulations in this range are perceived as particularly aversive (Li et al., 2018), cause faster behavioral reactions (Arnal et al., 2015; Ollivier et al., 2019), and increase neural responses in subcortical brain regions associated with aversive processing (Arnal et al., 2019).

The other features of screams (high intensity, broad spectrum, and high pitch) raise thorny measurement issues. Sound pressure levels cannot be measured directly from recorded audio files. High pitch can be gauged only with respect to a speaker's normative vocal register. Broad spectrum is evident in a wide range of vocalizations. Laughter, for example, can also feature high loudness, wide spectral range, and high relative pitch. It is the relative uniqueness and ease of measurement that makes roughness a useful operational measure, and coincidentally, a prime candidate for a universal cue signaling danger (Arnal et al., 2015).

To investigate whether scream-like music has the same roughness feature as, and is perceived similarly to, human screams, we conducted two studies. In the first study, we ran an acoustic analysis to test whether recorded screams and scream-like music exhibit enhanced roughness compared with control recordings. In the second study, we collected valence and arousal ratings for the audio files in order to test whether screams and scream-like music are perceived as sharing similar emotional qualities. We made the following hypotheses. First, we hypothesized that the mean power of the MPS within the roughness region (henceforth “roughness”) would be similar for screams and scream-like music, and would be significantly greater for screams compared to non-screaming vocalizations and for scream-like music compared to non-scream-like music. Second, given that roughness may be a universal cue for danger (Arnal et al., 2015), we hypothesized that roughness would correlate negatively with valence ratings and positively with arousal ratings for both music and vocal stimuli. Taken together, these results would demonstrate that scream-like music both sounds like and is perceived similarly to actual human screams.

The audio recordings used in the studies were deployed in a 2 × 2 factorial design, with one factor corresponding to the sound source (music, voice) and the other one to the scream-likeness of the sounds (scream-like, non-scream-like). Specifically, the four collections included (a) fearful scream vocalizations, (b) scream-like film music excerpts, (c) non-fearful human vocalizations (sounding similar to a held “ah” sound), and (d) non-scream-like film music excerpts as controls. All audio recordings are 800 ms in duration, RMS normalized, sampled at 16 kHz, and are in wav-file format.

In assembling a database of scream-like and non-scream-like music, we chose to make use of excerpts from horror film soundtracks. We chose the horror genre because descriptions of scream-like music that we found were typically describing horror movie soundtracks [i.e., Brown (1982)] and because horror films soundtracks are written with the explicit aim of scaring viewers [as opposed to violent or aggressive music, like death metal, which has been found to induce a range of positive and negative emotions in listeners (Thompson et al., 2019)]. Using an expertise-based approach, we curated excerpts from ten recently released films (2010 or later) that employed original composed soundtracks. Five scream-like music excerpts and five non-scream-like music excerpts were sampled from each of the ten horror movie soundtracks. In selecting potential scream-like passages, sampling focused on music written for especially terrifying scenes, such as moments of attack by a monster or ghost. On the other hand, non-scream-like music excerpts were pulled from scenes that were not especially terrifying and for which the music was deemed to be more emotionally neutral [see Table S1 for more information¹; to download the excerpts see Trevor et al. (2020)]. The curation process resulted in a database of 50 scream-like excerpts and 50 non-scream-like excerpts (Table S2 identifies the full list of excerpts and films¹).

Fearful screams and non-scream vocalizations were recorded from ten participants (five female, age: M = 27.20, SD = 3.39). Non-scream vocalizations consisted simply of a sustained “ah” vowel. Although the scream-like vocalizations were acted, past research has demonstrated that even acted screams have been judged to resemble real-life screams to a high degree (Engelberg and Gouzoules, 2018). Each participant produced ten audio recordings, resulting in a total of 50 fearful screams and 50 non-scream vocalizations.

Using matlab, the MPS of each excerpt was measured using the same procedure and equations as used by Arnal and colleagues (Arnal et al., 2015). Specifically, the initial spectrograms were obtained using a filter-bank approach with 128 Gaussian windows whose outputs were Hilbert transformed and then log-transformed. Then the modulation power spectra were obtained by applying a two-dimensional Fourier transform to the spectrogram (24 channels/octave) and log-transforming the resulting spectral power density estimates (Arnal et al., 2015). From there, the mean amplitude in the roughness range of 30 to 150 Hz along the temporal modulation range was taken [see Fig. 1(A); see Elliott and Theunissen (2009) for more detailed information on the MPS]. Results and statistical tests are reported after describing the behavioral study.

In the second study, 20 healthy participants (twelve female) reporting normal hearing and no psychiatric disorders participated in the rating experiment. Participants were recruited through the University of Zurich and received 15 CHF for their participation in the experiment. The participants were between the ages of 21 and 37 (M = 26.10, SD = 4.02). The procedure was approved by the Cantonal Ethics Commission of Zurich, Switzerland.

The experimental task for the rating experiment consisted of listening to each of the 200 audio files and rating the valence (from −3 to 3, with “−3” indicating the most negative valence and “3” indicating the most positive valence) and arousal (from 1 to 7, with “1” indicating lowest arousal and “7” indicating highest arousal) of the conveyed emotion using two analogical-categorical sliding scales. The experiment interface was created with matlab using Psychtoolbox. It took place in a quiet research room at the University of Zurich on a PC desktop computer using Sennheiser HD 200 headphones.

All statistical analyses were done using the r software package (version 3.6.1). We used the lm function to fit our standard linear regression models. For our mixed effect linear regression models, we used the “lme4” library (Bates et al., 2014) to fit the models and calculate t-values, and the “lmer” test package (Kuznetsova et al., 2017) to estimate p-values and degrees of freedom. The FDR adjusted p-values were calculated using the “p.adjust” function in r. Before model fitting, all continuous values were standardized and all categorical variables were coded as 0 and 1 (i.e., for sound type, music = 0 and voice = 1).

Recall that our first hypothesis predicted that scream-like music and screams would share a similar roughness level that would be higher than their matched controls. To test our first hypothesis, we used a standard general linear regression analysis. The predicted value was roughness and the predictor value was scream-likeness. We also tested for an interaction effect between sound types and scream-likeness. Finally, another standard general linear regression analysis was used to test for a main effect of scream-likeness on roughness for just the voice (coded as 1) for replicative comparison to the findings of Arnal et al. (2015). The results (reported in Table 1) demonstrate a significant main effect between scream-likeness and roughness driven by higher roughness values for scream-like stimuli as compared to non-scream-like stimuli across both sound types (p < 0.001); R² for the model was 0.407, and adjusted R² was 0.401. A similar effect was demonstrated in the regression analysis of the vocal stimuli only (p < 0.001) replicating the findings of Arnal et al. (2015); R² for the model was 0.826, and adjusted R² was 0.824. Additionally, the results showed a significant interaction effect between sound type and scream-likeness (p < 0.001) driven by a more extreme difference in mean roughness values between screaming voices and non-screaming voices compared to scream-like music versus non-scream-like music; R² for the model was 0.503, and adjusted R² was 0.495. These results are consistent with our hypotheses that roughness levels would be higher in the scream-like category than the non-scream-like category, across both sound types. However, contrary to our hypothesis, screams had a significantly higher mean roughness (M = 4.67, SD = 0.40) than scream-like music (M = 4.10, SD = 0.75, p < 0.001, see Table S3 for all unstandardized means and SDs¹).

Our second hypothesis was that roughness would correlate negatively with valence ratings and positively with arousal ratings for both music and vocal stimuli, supporting its reputation as an aural cue for danger (Arnal et al., 2015). To test this hypothesis, we used emotion ratings (valence and arousal) as the predicted values for two mixed effects linear regression models. The predictor value was roughness. Once again, participant was included as a random slope. We also tested for a main effect for roughness on emotion ratings for vocal stimuli only for further replicative comparison to the findings of Arnal et al. (2015). Additionally, we tested for an interaction effect between sound type (music vs voice) and roughness. The regression analyses results are reported in Table 2. Consistent with our hypothesis, roughness correlated negatively with valence ratings for both musical stimuli [β = −0.196, SE = 0.032, t = −6.15, p < 10⁻⁴, BH-adjusted p < 10⁻⁴] and vocal stimuli [β = −0.443, SE = 0.052, t = −8.599, p < 10⁻⁸, BH-adjusted p < 10⁻⁷]. Also consistent with our third hypothesis, roughness correlated positively with arousal ratings for both musical stimuli [β = 0.07, SE = 0.025, t = 2.859, p = 0.01, BH-adjusted p = 0.0134] and vocal stimuli [β = 0.404, SE = 0.057, t = 7.144, p < 10⁻⁷, BH-adjusted p < 10⁻⁶]. Interestingly, there were significant interaction effects between roughness and sound type for both valence [β = −0.247, SE = 0.045, t = −5.441, p < 10⁻⁵, BH-adjusted p < 10⁻⁵] and arousal ratings [β = 0.334, SE = 0.058, t = 5.746, p < 10⁻⁵, BH-adjusted p < 10⁻⁵]. Both regression slopes were steeper for the vocal sound type than for the musical sound type. The interaction effect models can be seen in the scatterplots in Fig. 1(B). In spite of the significant interaction effect, the significant main effects for both valence [β = −0.330, SE = 0.037, t = −8.842, p < 10⁻⁸, BH-adjusted p < 10⁻⁷] and arousal [β = 0.251, SE = 0.034, t = 7.342, p < 10⁻⁷, BH-adjusted p < 10⁻⁶] indicate that the relationship between roughness and emotion ratings extends across sound types, consistent with our hypothesis.

This research was inspired by the frequent comparison of scary film music to human screams (Brown, 1982). Our motivating question was whether scary film music mimics an acoustic feature unique to human screams (roughness) (Arnal et al., 2015) in order to scare the viewers. In order to address this question, we calculated roughness levels (specifically, the average MPS power in the 30–150 Hz region) (Arnal et al., 2015) for four groups of audio recordings: screams, non-screaming vocalizations, scream-like music, and non-scream-like music. We also ran a behavioral study where 20 participants rated the arousal and valence of each audio file.

Consistent with our hypotheses, we found that both screams and scream-like music exhibited a higher level of roughness and were rated as having a more negative valence and a higher arousal level than their non-screaming counterparts. However, contrary to our hypotheses, screams had a higher roughness level than scream-like music. Overall, the results demonstrated a greater difference in roughness levels and emotion ratings between the vocal stimuli than between the musical stimuli. These results suggest that while scream-like music does seem to sound like and be perceived similarly to human screams, the musical rendition is still a muted version of the real thing and therefore may not provoke as potent of a reaction. This finding is notably in opposition to the super-expressive voice theory (Juslin and Västfjäll, 2008) that music is capable of amplifying vocal affective behaviors beyond the capability of the vocal system. Perhaps screams are an exception to this theory. Overall, the results suggest that roughness can effectively translate from a vocal cue for danger into a musical cue for danger. It is therefore reasonable to suggest that scream-like music might scare viewers in part because it is evocative of a human scream, a naturally alarming sound.

It is important to note that our results may have been biased by the non-blind selection of the film music excerpts. This work corroborates the findings of several recent papers investigating roughness in music (Arnal et al., 2015; Belin and Zatorre, 2015; Blumstein et al., 2010; Blumstein et al., 2012; Liuni et al., 2020; Ollivier et al., 2019) while being the first to use the MPS to investigate the presence of roughness in scary film music. Future work might test similar hypotheses by directly manipulating roughness cues in music [i.e., Anikin (2020)], by investigating other features associated with vocal roughness such as jitter, shimmer, and HNR (Liuni et al., 2020), or by investigating whether our findings could be extended to other music found to induce negative emotions, such as metal music for non-fans.

C.T. received funding from the European Union's Horizon 2020 research and innovation program under the Marie Skłodowska-Curie Grant Agreement (No. 835682). S.F. received funding from Swiss National Science Foundation (Grants Nos. SNSF PP00P1_157409/1 and PP00P1_183711/1). The authors thank Lawrence Feth for guidance regarding the acoustic analyses and to David Huron for valuable feedback on the project. Finally, the authors thank Arkady Konovalov for helpful input regarding the statistical analyses.