Screaming, Yelling, Whining and Crying: Categorical and intensity differences in Vocal Expressions of Anger and Sadness in Children's Tantrums

Abstract

Young children's temper tantrums offer a unique window into the expression and regulation of strong emotions. Previous work, largely based on parental report, suggests that two emotions, anger and sadness, have different behavioral manifestations and different time courses within tantrums. Individual motor and vocal behaviors, reported by parents, have been interpreted as representing different levels of intensity within each emotion category. The present study used high fidelity audio recordings to capture the acoustic features of children's vocalizations during tantrums. Results indicated that perceptually categorized screaming, yelling, crying, whining, and fussing each have distinct acoustic features. Screaming and yelling form a group with similar acoustic features while crying, whining, and fussing form a second acoustically related group. Within these groups, screaming may reflect a higher intensity of anger than yelling while fussing, whining and crying may reflect an increasing intensity of sadness.

Temper tantrums are both prevalent and frequent in young children, and they often present a serious management problem for their parents. Tantrums are challenging to prevent and even more difficult to interrupt once underway. Their emotional momentum speaks to the intensity of the emotions involved. Furthermore, tantrums that are excessive in early childhood and/or persist into later childhood can predict future maladjustment and even psychopathology (Caspi, Elder & Bem, 1987, Stevenson & Goodman 2001, Stoolmiller 2001). Because tantrums offer a window onto the early expression and (dys)regulation of strong emotions that are otherwise difficult to observe, they are a compelling phenomenon for scientific study (Potegal & Davidson, 2003.)

Goodenough's (1931) pioneering monograph Anger in Young Children was an early landmark which drew from parental diaries to chronicle outbursts of 45 children between 6 months and 8 years of age. The most frequent motor acts during tantrums were kicking, stamping, jumping up and down, hitting, and throwing the self on floor, which occurred in somewhere between 3% and 28% of tantrums. The most common vocal behaviors were classified as cry, scream, fuss, whine, and snarl, occurring from 25% to 85% of tantrums, depending on age. Analyzing a new set of narratives written more than 60 years later by the parents of 335 children 18 months to 5 years old, Potegal & Davidson (2003) found that crying was the most frequent vocal expression, occurring in 86% of tantrums. Screaming and shouting (hereafter termed yelling) occurred in about 40% of tantrums, and whining in about 13%.

The latter prevalence estimates are remarkably similar to those of Goodenough (1931). In our view, however, anger is only half the story of tantrums. Factor analyses by Potegal & Davison (2003) indicated that these motor and vocal acts form groups that represent varying degrees of two different emotions, anger and “distress” (sadness and comfort-seeking). With regard to anger, Potegal & Davidson (2003) found three factors that could be ranked as different in levels of intensity. Scream loaded with kick, hit, and stiffen on a factor labeled High Anger. Yell loaded with throw on a second factor labeled Intermediate Anger. The lowest intensity anger factor, defined by stamping, contained no vocal component. Multidimensional scaling showed that all three factors had a similar temporal pattern of peaking near the onset of the tantrum and declining thereafter suggesting that they were indeed expressions of the same emotion (Potegal, Kosorok, & Davidson, 2003). The specific behavioral components of the respective factors provided face validity for their identification as different intensities of anger. This identification was further supported by the progressively higher correlations of Low, Intermediate, and High Anger with tantrum duration and autonomic activation as well as with parental judgments of overall tantrum intensity.

“Distress” in these publications referred to expressions of sadness and cooccurring comfort- seeking. The vocalizations whine and cry loaded together on a factor representing sadness. These vocalizations were more evenly distributed across the tantrum in a temporal pattern distinctly different from the anger-related behaviors. Anger and sadness factors were also differentially correlated with likelihood of parental intervention during the tantrum. Because the present report does not address comfort-seeking, we will refer to the second tantrum emotion simply as sadness. However, the use of anger and sadness as descriptors of emotions displayed during tantrum behaviors should not be interpreted as implying continuity with adult expressions of anger and sadness¹.

The model of tantrums as composed of groups of behaviors reflecting anger and sadness at different levels of intensity (Potegal & Davidson, 2003) has been extended and confirmed by the finding that the on-ward “rages” of older child psychiatry inpatients have a structure quite similar to the tantrums of younger, non-psychiatrically disturbed children (Potegal, Carlson, Margulies, Gutkovitch, & Wall, 2009). Interestingly, sadness appeared at two levels of intensity in the older children's data.

The observation that different types of vocalizations are important in distinguishing anger vs. sadness during children's temper tantrums suggests that a more detailed analysis of their acoustic characteristics might further clarify differences between tantrum emotion expressions as well as between different intensities of the same emotion. In adult speech, several studies support the acoustic differentiation of different emotions. Banse and Scherer (1996) reported that adult actors’ portrayals of anger (both mild and severe) versus sadness (both mild and severe) differed on several acoustic features (see Banse & Scherer, 1996, Table 6). Sadness was associated with longer vocalizations and vocalizations with lower concentrations of energy in the 0-1 kHz range. Sad speech also had less energy overall, as well as somewhat lower fundamental frequencies. Hot anger, in contrast, had a higher fundamental frequency, more energy overall, and a greater concentration of energy above 1000 Hz. Juslin and Laukka (2001, Table 8), using similar methods, reported acoustic characteristics of angry and sad speech that were consistent with these characterizations. Indeed, studies of infant crying have examined similar acoustic features and found that fundamental frequency, duration, and energy distribution are related to how aroused the infant is presumed to be; that is, how much hunger or pain is being experienced (see Green, Gustafson, & McGhie, 1998; Porter, Porges, & Marshall, 1988).

Table 6.

Standardized Canonical Discriminant Function Coefficients

	1	2	3	4
Duration	.050	.823	-.259	.451
Total Energy	-.741	-.338	.088	-.100
Peak Frequency	-.061	.041	.034	.214
F0 Mean	.170	.331	.421	-.266
F0 SD	-.235	.086	-.278	.100
Energy Band1	1.091	.101	1.993	.901
Energy Band2	1.199	.449	2.059	.395
Energy Band3	.453	.372	1.920	.502
Energy Band4	-.059	-.160	.308	.525
Energy Band5	.135	.272	.279	-.610

Note. Entries in bold are the coefficients used to interpret the 4 discriminant functions (see text for details). These 4 combinations of acoustic features were able to correctly classify 38% of Fusses, 87% of Whines, 26% of Cries, 46% of Yells, and 79% of screams.

However, the studies on adult vocal expressions have all used actors to portray the emotions, and there is a compelling need for naturalistic samples of vocal expressions of emotion (see review by Green, Gustafson, & Whitney, 2010). As noted above, children's tantrums offer abundant examples of spontaneous emotion-related vocalizations. However, previous coding of tantrum vocalizations has been by parents at home, or staff of a psychiatric inpatient facility, and their judgments were made while trying to manage the tantrums as well as to record them. These concerns raise a series of methodological questions. Can well-trained observers perceptually classify vocalizations reliably using the above mentioned categories (i.e., scream, yell, cry, whine)? If they can, what are the acoustic properties that enable them to do so (i.e., what are the bases of their perceptual classifications)? Do the acoustic properties that distinguish anger from sadness vocalizations during children's temper tantrums support the distinction between these emotions and/or provide clues about how they differ? Finally, do different types of vocalizations suggest different intensities of anger or sadness?

To answer these questions, we used new technology to collect the high quality audio recordings necessary for acoustic analysis. To the best of our knowledge, this is the first study to analyze high fidelity audio recordings of emotion-related vocalizations in a naturalistic situation. Our goal was to define the acoustic characteristics of tantrum vocalizations and to test the following specific hypotheses:

1. The majority of tantrum vocalizations can be classified as scream, yell, cry, and whine by trained observers

2. There is a basis for these perceptual classifications in specific acoustic characteristics that are similar to those used in previous studies of infants’ and adults’ vocal expressions of emotions.

3. Vocal expressions of anger and sadness can be further divided into vocalizations representing different emotion intensities. This proposal can be sorted into four subhypotheses:

   1. Among the vocalization types, scream and yell are acoustically most similar to each other
   2. Cry and whine are acoustically most similar to each other
   3. Referring to previous characterization of adult emotional speech, yell, and scream are more anger-like, whine and cry are more sadness-like
   4. Within their respective emotion categories, scream represents a higher intensity of anger than yell, cry represents a higher intensity of sadness than whine

Method

Participants

Participants were 13 two and three year old children (M_age = 31 mos, SD = 6 mos)) whose families lived in the Minneapolis-St. Paul MN area. Seven were boys. Two of the children were Hispanic, the others were Caucasian. More than 80% of fathers and 90% of mothers had Bachelors or Masters degrees. Median family income was $75,000-$85,000. Children were recruited from the Infant Participant Pool of volunteers maintained by University of Minnesota's Institute for Child Development by a 3-step, IRB-approved consent process: 1) After being contacted by a member of the research team, parents who consented to a telephone interview reported the frequency and average duration of their child's tantrums, 2) those with frequencies ≥ 3-6/week and durations ≥ 1-2 min were re-contacted by one of the authors (MP), who explained the purpose of the research, manipulations to be performed, measures to be taken, and compensation, and 3) interested parents were sent a written consent form. Those who consented were enrolled. Exclusions were children with any major developmental, language, or physical health disorder.

Recordings

Audio recordings were made using professional equipment specially chosen for the project (equipment selection, assembly and testing by Orfield Laboratories, Minneapolis MN). A Countryman EMW microphone and Lectrosonics MM400A transmitter were contained in a pouch sewn to the front of a toddler sized “onesie.” A Lectronics R400A receiver and Marantz PMD-670 digital audio recorder were housed in a cabinet, and the system was calibrated in the laboratory. Because the PMD-670 omits the range-compression typical of consumer products, the system was found to have a flat frequency response up to 4 kHz in the range of 35 to 115 dBA. (The sampling rate was 48kHz, with 16-bit encoding.)

The audio system (and an accompanying video system) were set-up in a room in the child's home during a 2 hour installation and parent training session. On subsequent day(s) and time(s) of her choosing, the parent inserted a battery in the transmitter, dressed the child in the onesie, and activated the system. Recording then took place continuously for up to 4 hours at a time (limited by life of the transmitter battery.). When a tantrum occurred, the parent turned off the system at its end; a research assistant then returned to the home to retrieve the recording.

Coding

The digital files representing each tantrum were coded using Praat, a program designed for acoustic analysis of speech (Boersma, 2001). The basic unit of coding was the child's vocalization, defined as any audible activity of the vocal tract emitted during the course of a single respiratory expiration. (The intake of air during the inspiration phase was typically audible on the recordings and made segmentation of each vocalization relatively simple.) The first phase of the acoustic analysis, then, involved marking the beginning and end of each vocalization on the digital file. There were a total of 2,543 vocalizations during the 24 tantrums.

The second phase of the analysis involved coding each vocalization into one of 5 categories (Table 1), based on the tantrum studies reviewed previously and also in part on previous studies of crying during the first 18 months of life (e.g., Green et al. 1988). The previous tantrum work categorized vocalizations as scream, yell (or shout), cry, or whine, whereas the previous infant cry work distinguished cries and fusses. Careful analysis of the audiotaped tantrums suggested a class of vocalizations different from whining and crying and similar to infant fusses, so this fifth category was added to the intended 4-way classification. Intercoder agreement at this stage was based on two coders independently scoring 5 tantrums for a total of 349 vocalizations. Agreement was high, with 93% agreement and kappa = .91. Digital spectrograms of exemplars of each type are shown in Fig. 1.

Table 1.

Definitions of Vocal Expressions Coded During Temper Tantrums

Scream	Typically shrill, loud, and with no verbal content. Usually short and flat melody.
Yell	Typically short in duration, command-like and usually containing some verbal content. Loud, but not as shrill as a scream.
Cry	Relatively loud and effortful, typically with and up and down melody. Breath may be interrupted, as in sobbing. Similar to an infant's cry.
Whine	Typically contains some verbal content with an up and down melody. May also include relatively shrill, monotonous nonverbal vocalization.
Fuss	Typically short, flat or falling melody, relatively quiet and low pitched.

Figure 1.

Sample spectrograms of fuss, whine, and cry, (left column, reading down) and yell and scream (right column, reading down). Duration is on the X axis (and varies by type of sound) and frequency is on the Y-axis (from 0 to 5000 Hz with grid lines every 1000 Hz). Darkness of lines indicated more energy in that frequency range. See text for description of acoustic characteristics of each type of vocalization.

Of the 2,543 vocalizations, 1,022 co-occurred with parent speech or other ambient noise. Of the remaining high quality vocalizations of the child alone, about 7% were not clearly emotion-related sounds or were otherwise uncodable, and about 7% did not fall clearly into one of the five categories, typically because they exhibited characteristics of more than 1 of the 5 categories. The final corpus of sounds used in the acoustic analysis totaled 1,299 individual vocalizations. The frequency distribution of these sounds across children and tantrum episodes is given in Table 2. It should be noted that hypothesis 1 was clearly supported; of the vocalizations these children emitted during tantrums, almost 85% could be classified into the five categories in Table 1.

Table 2.

Frequency Distribution of Vocalization Categories Used for Acoustic Analysis

Tantrum Number	Fuss	Whine	Cry	Yell	Scream
1	159	357	115	15	9
2	64	233	71	99	63
3	22	25	64	00	03
Grand Total	245	615	250	114	75

Note. Seven children were recorded for two tantrums each, four for one tantrum, and two children contributed three tantrums.

Each vocalization was then analyzed for 11 different acoustic parameters (Table 3). These parameters were chosen because they are similar to those used in analyzing adults’ vocal expressions of anger (see review in Green et al. 2010) as well as those used in characterizing infants’ cry sounds (e.g., Gustason, Green, & Tomic, 1984; Green, Jones, & Gustafson, 1987). The fundamental frequency² of a vocalization has been especially prominent in both literatures, with higher fundamental frequencies (generally perceived as higher pitch) associated with higher distress, but overall duration, manner of phonation (related to harmonicity), and distribution of energy in different frequency bands have been important in perception of distress as well.

Table 3.

Definitions of Acoustic Measures

Duration	Time from beginning to end of each expiratory segment (in msec).
Fundamental Frequency (F0)	Mean and SD of the extracted pitch trace for the entire vocalization frequency using an autocorrelation estimation method. Pitch extraction range set at 200-700 Hz.
Energy	Sum of the squared amplitudes of the time sampled vocalization
Harmonicity	The degree of acoustic periodicity, expressed in dB.
Peak Frequency	The frequency in the Long Term Average Spectrum with the greatest amplitude (in Hz)
Spectral Energy	The energy in selected frequency bands from the overall spectrum of the vocalizations. Expressed as a proportion of the total energy. Bands: 100-500 Hz, 500-1000 Hz, 1000-1500 Hz, 1500-2000 Hz, and 2000-2500 Hz.

Note. Further description of these measures is available from the authors or from the manual available at www.praat.org.

Analyses

Because the design of the study involved nesting of vocalizations within tantrums, which were nested within participants, multilevel modeling was employed to compare the vocalization categories (i.e., fusses, whines, cries, yells, and screams) on each of the 11 acoustic features. In addition, discriminant function analysis was used to describe the optimal linear combination of acoustic features for separating the vocalization categories.

Results

Differences Among Vocalization Types: Preliminary Comparisons

Simple one factor ANOVAs indicated that the five perceptually-based categories of tantrum vocalizations were significantly different in mean values for every one of the acoustic features. Fusses were shortest in duration (Table 4), with the least overall energy, lowest fundamental frequency and peak frequency, and least energy above 1000 Hz (frequency bands 3, 4, and 5). The distribution of energy for whines was also concentrated below 1000 Hz, but whines were longer in duration with slightly higher fundamental frequencies and peak frequencies. Cries were the longest sounds (M= 1.87 sec), with greater structure in the harmonics, high fundamental frequencies (M = 397.43) and more energy in the 500-1500Hz range. Screams contained the most energy, the highest peak frequency (M = 1297 Hz), the largest variation in fundamental frequency, and the most energy in the 1500-2000 Hz frequency band (band 4). Clearly, the perceptual categories of tantrum vocalizations have many acoustic bases, supporting hypothesis 2.

Table 4.

Means, Standard Deviations and AVOVA results for Five Sound Types

	Sound Type
	Fuss	Whine	Cry	Yell	Scream	F(4,1294	np2
Duration	046wcys (.39)	1.38 fcys (1.00)	1.87fw (1.22)	1.64fw (.84)	1.66fw (.65)	83.17***	0.21
Energy	0.000cys (.001)	0.002cys (.004)	0.006fwys (.007)	0.017 fwcs (.015)	0.034fwcy (.021)	355.28***	0.52
Harmonicity	5.89wcys (9.04)	12.50 fcys (10.01)	13.44fws (6.83)	14.51fws (4.39)	6.66fwcy (4.16)	39.66***	0.22
Peak Freq	609.59wcys (220.19)	693.74fcys (233.70)	858.00fwys (291.40)	1028.95 fwcs (325.46)	1296.67 fwcy (449.12)	139.39***	0.30
Mean F0	361.92cy (86.59)	376.10cy (55.28)	397.43fwys (51.97)	426.22 fwcs (51.23)	377.60cy (65.49)	26.18***	0.07
SD of F0	50.75wcys (35.23)	61.40fcys (28.91)	72.44fws (26.42)	79.11fws (27.33)	102.16fwcy (30.57)	55.91***	0.15
Band1	0.35wcys (.28)	0.28fcys (.27)	0.12fws (.13)	0.08fw (.09)	0.02fwc (.03)	66.90***	0.17
Band2	0.56ys (.26)	0.57ys (.26)	0.61ys (.23)	0.40fwcs (.24)	0.22fwcy (.22)	48.10***	0.13
Band3	0.05wcys (.09)	0.09fcys (.14)	0.13fwys (.16)	0.35fwc (.22)	0.35fwc (.18)	141.20***	0.30
Band4	0.02cys (.04)	0.03ys (.05)	0.05fys (.07)	0.08fwcs (.10)	0.20fwcy (.17)	116.55***	0.27
Band5	0.02cys (.03)	0.02cys (.03)	0.06fws (.07)	0.06fws (.06)	0.11fwcy (.10)	78.60***	0.20

^***p<.001

Note. See Table 3 for basic definitions of these acoustic features. Superscripts next to a mean indicate which other groups are significantly different from the given mean.

Multilevel Modeling

A more appropriate analysis of differences across the vocalization categories, however, takes into account the nested structure of the data (Table 5.) Multilevel modeling proceeded by fitting two models for each of the 11 acoustic features. First, a base model was fit that predicted the acoustic feature of each vocalization (level 1) from parameters reflecting the mean for each participant (the fully independent unit in the study, level 3) and for each tantrum nested within participant (level 2). A second model was then fit that added a level one predictor (i.e., vocalization category) to the base model. The significance of the change in fit was evaluated for each acoustic feature and was significant in every case (Table 5). The intraclass correlations for the acoustic features ranged widely, from .03 to .21, indicating that most acoustic features showed a moderate proportion of variance attributable to infants. Overall, there was striking agreement between the standard ANOVAs (Table 4) and the more appropriate multilevel modeling of acoustic features (Table 5).

Table 5.

Multilevel Model Results

	Estimated Marginal Means
	Fuss	Whine	Cry	Yell	Scream	-2LL Chg	ICC	F
Duration	0.57wcys	1.31fcys	2.04fw	1.98fw	2.00fw	337.56***	0.19	108.82***
Energy	0.00cys	0.00cys	0.01fwys	0.01fwcs	0.03fwcy	642.50***	0.08	177.93***
Harm.	6.95wcy	12.85fs	14.56fs	15.60fs	7.68wcy	137.07***	0.05	116.22***
Pk. Frq.a	613.95wcys	701.24fcys	849.43fwys	989.60fwcs	1252.60fwcy	338.19***	0.03	549.86***
Mean F0	386.38cy	393.17cy	413.64fws	428.00fws	379.17cy	68.46***	0.21	712.23***
SD F0a	55.05wcys	61.94fcys	74.43fws	76.99fws	99.17fwcy	147.94***	0.06	186.20***
Band1a	0.36wcys	0.29fcys	0.14fw	0.13fw	0.08fw	173.54***	0.06	73.99***
Band2a	0.54ys	0.53cys	0.59wys	0.36fwcs	0.18fwcy	161.36***	0.10	128.08***
Band3	0.05wcys	0.10fcys	0.14fwys	0.29fwc	0.30fwc	244.38***	0.12	70.11***
Band4	0.03wcys	0.04fys	0.05fys	0.10fwcs	0.21fwcy	368.64***	0.15	94.11***
Band5	0.02cys	0.02cys	0.05fws	0.06fws	0.11fwcy	199.26***	0.12	51.52***

^aBecause of small sample sizes at levels 2 and 3 and distributional properties of these variables, only 2 tantrums could be nested at level 2 under the level 3 individual participants.

Superscripts next to a mean indicate which other groups are significantly different from the given mean.

Multivariate Analyses

To supplement the univariate analyses above, take into account covariance among acoustic parameters, and evaluate hypothesis 3, two discriminant function analyses were employed to determine the number and nature of combinations of acoustic features that separated vocalization categories. For these analyses, harmonicity was dropped because this it was undefined for 157 of the vocalizations Two sounds were missing the measure of variability of the F₀, and these were eliminated from the discriminant analysis, leaving a total of 1297 vocalizations. First, the presumptive anger and sadness vocalizations were combined to perform a 2-group discriminant analysis. Scream and yell were combined and compared with cry, whine, and fuss. In this analysis, the single discriminant function was highly significant, canonical r = .73, and Wilks’ lambda = .466, p < .001. The highest loading variable on the standardized discriminant function was .725 for energy, and the other high loadings in absolute value (-.354, -.443) were for the proportion of energy in the two lowest frequency bands, 100-500Hz, and 500-1000Hz. Thus, yells and screams had much more energy overall and the fusses, whines and cries had proportionately more energy in the lower frequency ranges. Figure 2 shows discriminant scores for the high versus low intensity sounds, and it is evident that very little overlap was found.

Figure 2.

Distribution of scores for low intensity (fuss and whine) and high intensity sounds (cry, yell, and scream) on the discriminant function. Note that there is almost no overlap in scores, and the classification accuracy is 93%. See text for a description of the variables loading on the discriminant function.

The prediction of group membership was excellent, with 92% of the sounds correctly classified as one or the other group (98% of fuss, whine, and cry sounds were correctly classified, along with 65% of yell and scream sounds). All other combinations of vocalizations into 2 groups (e.g., scream and cry vs. yell, whine and fuss) yielded substantially fewer correct classifications on discriminant analysis, ranging from 64% to 78%. This comparison suggests that yell and scream vs. fuss, whine and cry is the most appropriate categorical grouping³

A second discriminant analysis was performed using all 5 vocalization categories. Here, all four discriminant functions were statistically significant and accounted for 81%, 13%, 4%, and 2% of the discriminating power, respectively (Table 6). The canonical correlations for the functions were .814, .488, .312, and .227. Examination of the standardized coefficients showed that vocalizations that were relatively low in total energy (that is, were quiet) and had their energy concentrated in bands below 1000 Hz had high scores on Function 1. Indeed, fusses and whines generally fit this profile and have high group means on Function 1 (Figure 3).

Figure 3.

Group centroids for fuss, whine, cry, yell, and scream in discriminant 3-space. See text for descriptions of the 3 discriminant functions.

Function 2 separated vocalizations long in duration and with relatively high energy in the 500-1000 Hz band from the others. The clear separation here was of Scream and Fuss (i.e., relatively short, little energy in the 500-1000Hz range) versus Cry and Yell. Function 3 contrasted sounds with relatively large amounts of energy less than 1500 Hz (i.e., frequency bands 1, 2, and 3) from other sounds. Yells scored highly on this third function. However, the 5 group discriminant analysis (which used prior probabilities determined by the relative frequency of the sound types) correctly classified only 806 of the 1297 sounds (62%) according to their acoustic characteristics. This result would be expected if differences within the presumptive categorical classifications noted above were matters of intensity, in which there might well be notable overlaps. Fig. 4 shows exactly this situation of substantial overlap among fuss, whine and cry in the discriminant plane.

Figure 4.

Distributions of fuss, whine, cry, yell, and scream (10th to 90^th percentiles encircled) in the discriminant plane of Function 1 and Function 3.

Taken together, the 2 discriminant analyses support hypotheses 3A and B: That is, there is better than 90% correct prediction of the presumptive anger (scream, yell) versus presumptive sadness (cry, whine, and fuss) categories. With regard to hypothesis 3D, the five vocalizations are distributed across Function 1 in the order scream>yell>cry>whine> fuss, meaning that these vocalizations had progressively less energy overall and less energy above 1000 Hz, suggesting that within-category intensity may be related to energy at less than 1500 Hz.

A final statistical analysis compared the ‘melody’ or change in fundamental frequency over time for each vocalization type (Figure 5). To examine relative change in F₀, the mean F₀ was subtracted from the F₀ value at 10 equally-spaced segments within each vocalization. In this transformation, shown in Fig. 4, yell and scream each had an early, sharp peak while cry and whine showed later, much more rounded peaks. A 5 (vocalization category) by 10 (position in the expiratory segment) mixed ANOVA was computed, with a special contrast to compare scream plus yell with the average of cry, whine, and fuss. (The contrast coefficients were .5, .5, -.333, -.333, and -.333 for scream, yell, cry, whine, and fuss, respectively.) The interaction of vocalization category and time into sound was significant, F(36, 10566) = 3.582, p < .001, and the contrast of (scream + yell) vs. (cry + whine + fuss) was significant, p < .001. As in the 5 group discriminant analysis, then, scream and yell were similar to each other but different from cry, whine, and fuss. The qualitative similarities and differences in the shapes of the curves and the quantitative results of the ANOVA support the putative categorical distinction between yell and scream vs. whine and cry.

Figure 5.

Melody contours for each type of vocalization. F₀ was measured at 10 equally-spaced points in each sound, and each value in each sound was subtracted from the mean F₀ for that sound.

Finally, to place these sounds in the context of other negative vocalizations, per hypothesis 3C, we compared F₀ values to published data on infant crying. Many publications exist on the acoustic features of newborn cry sounds, but there are relatively few studies of acoustic features from later ages. Wermke, Mende, Manfredi, and Bruscaglioni (2002) reported longitudinal data on 6 infants showing little change in F₀ during the first 6 months, with values ranging from 377 to 588 Hz. Baeck & Nogueira de Souza (2007) reported a range of 385 to 410 Hz from 30 infants in the first 6 months of life.

Perhaps the most relevant data come from Rothgänger (2003), who reported an average F₀ of 463 to 480 Hz for the cry sounds of 15 infants across the entire first 12 months, which was considerably higher than the 342 Hz for babbling sounds during the same time period. However, the present value of almost 400 Hz for the F₀ of the cry sounds (Table 4) is similar to the values for cry sounds reported by all 3 studies.

Discussion

Previous empirical research on young children's temper tantrums has yielded a model proposing two sets of motor and vocal behaviors, one that expresses anger and the other, sadness (Potegal & Davidson, 2003; Potegal et al 2009). Within this model, four distinct types of tantrum vocalizations were identified for younger children, with whine and cry associated with a factor labeled sadness while scream and yell (previously termed shout) were associated with a factor labeled anger. However, the previous research on tantrum vocalizations was based on parent report, not on actual audio recordings. In the present work, we evaluated the previously proposed distinctions among tantrum vocalizations using novel data collection methods in naturalistic situations. We found that the majority of tantrum vocalizations could be reliably perceptually classified as scream, yell, cry, and whine. We also identified a 5^th, low energy type that we called fuss based on previous research on infant crying. Both the perceptual and the statistical reliability of these identifications were high. Together, the five types accounted for almost 85% of tantrum vocalizations.

Furthermore, young children's tantrum vocalizations identified perceptually are readily separable by their acoustic features. These acoustic features overlap considerably with those previously identified in analyses of both adult and infant expressions of anger and sadness. Banse & Scherer (1996), and others, have reported that adults’ angry speech (both ‘hot’ and ‘cold’ angry speech) is relatively short, has more energy overall, higher fundamental frequencies, and more energy in high frequencies. Results of our analysis of tantrum vocalizations are in substantial agreement with the findings on adult emotion speech. Screams and yells had more overall energy, higher fundamental frequency (especially yells), and more energy above 1000 Hz. The only contrary finding is in the duration of the sound; yell and screams were relatively longer compared with whines, although a bit shorter than cries.

One difficulty in making comparisons between the vocalizations in the present study and those used in adult studies is the source of the sounds. Adult studies have invariably used staged emotion speech, sometimes even the same phrase spoken in anger versus sadness. Thus, comparison with the present, naturalistically recorded vocalizations, especially on duration measures, is problematic. In addition to the staged speech, one other factor in our naturalistic recordings it that expressions of sadness during preschoolers’ temper tantrums are influenced by juxtaposed expressions of anger. This potential carryover effect would not be present in adults’ portrayals of sadness in the experimental literature on vocal expressions of emotions.

Fuss sounds had the least overall energy, the lowest peak frequency, and the least energy above 1 kHz. These low energy, low frequency distress sounds have been studied extensively during the period of infancy as part of the literature on crying. There, fusses are recognized as briefer, less intense negative sounds than full blown wails (Gustafson et al., 984). Given that fusses are located near to whines in discriminant space (Figs 3 and 4), it is perhaps understandable that neither sound is especially salient to parents during tantrums. In fact, although Potegal and Davidson (2003) previously coded whining as an indicator of distress, they did not find justification for using it alone as a marker for the onset of a temper tantrum. Whether future studies should continue to separate whine and fuss would depend on the purpose of the study, as both seem to be indicators of low intensity distress.

One of our hypotheses was that a pattern of similarities and differences among acoustic characteristics would emerge such that the several types of tantrum vocalizations could be effectively classified into two groups; further, these groups would be consistent with the emotions of anger and sadness. We found that the best discrimination in the classification analysis, amounting to correct identification of over 90% of cases, grouped scream and yell together and cry, whine, and fuss together. Complementing this result was a pattern of qualitative similarities and differences in “melody.” Yell and scream both had early, sharp peaks in their fundamental frequency within vocalizations over time while cry and whine showed later, much more rounded peaks (Fig. 5.) The acoustic distinction between yell/scream vs. fuss/whine/cry parallels the distinction between the prosody of angry vs. sad utterances by adults.

There was also converging evidence for graded acoustic differences within emotion groupings. Both in terms of energy distributions across frequency bands and temporal contours of F₀ (Figure 5), screams could be viewed as more intense than yells. Similarly, fuss, whine and cry appeared to be ordered by increasing intensity according to both energy distributions across frequency bands and F₀ temporal contours. The gradation of intensity within a class of vocalizations is not unlike what has been proposed for the production of infant cries (Gustafson, Wood, & Green, 2000); in that literature, cry sounds are generally perceived as more intense than fuss sounds but both are classified as ‘negative’ vocalizations.

One theoretical controversy in the area of emotion expressions centers on whether emotional expressions represent discrete or continuous processes. Some argue vigorously for separable physiological states and behavioral (especially facial) manifestations of happiness, sadness, anger, disgust, pride, and so on. Others argue that two underlying dimensions of arousal and valence are responsible for the various expressions of emotion (Russell, Bachorowski, & Fernández-Dols, 2003; Barrett, 2006). Although we have presented arguments favoring the categorical, discrete emotions view, our empirical base is continuous acoustic features that were used to describe categories of sounds reliably coded by trained listeners. It is worth noting here that categorical perception is, in fact, a rather ubiquitous process across stimulus domains (including spoken language) and indeed across species (see review by Hauser, 2001). More relevant data for the discrete versus continuous arguments in the emotion expression literature might have to come from combining perceptual, behavioral, and functional neural architecture data.

Of course, there are several limitations of this research. First, the sample is a high SES group which was preselected for higher tantrum frequency and duration. However, the vocalizations and behaviors during these tantrums are similar to those reported as far back as Goodenough (1931), so it seems unlikely that these tantrums are different in quality. Second, the sample size is relatively small. This limitation is due to the dense, naturalistic behavior samples required for acoustic analysis. Finally, the acoustic features themselves are a limited set. Although the features are similar to those studies in adult emotion expressions as well as in infant distress studies, they do not contain many features used in, for example, language processing Future studies might expand this set and focus on formant transitions or other features adults use in processing language.

Although the present study did not analyze the behaviors of the children, the co-occurrence of vocal expressions with behavioral indicators, such as kicking, hitting, and dropping down, should be one of the next steps in this research. Whether behavioral indicators, such as hit or kick, and vocal indicators, such as scream, can be substituted for one another as manifestations of high anger, or whether these expressions necessarily co-occur or synergize each other, is unknown. These questions are presently being studied in our laboratories, and the answers will bear on the important question of how young children express and regulate intense emotions. Temper tantrums are ideal, naturally-occurring, episodes to examine issues of emotion reactivity and regulation.