That’s How You Write a Song
An Analysis of Music Characteristics on Popularity at the
Eurovision Song Contest
Noé Martin 12180041 Master Thesis Graduate School of Communication Master’s programme Communication Science Supervisor: Jeroen Lemmens 26/06/2020
Abstract
Understanding which musical features that foster popularity in songs is essential for both individual artists and the music industry at large. Predictors of music popularity have been studied in relation to charts and rankings but not in the context of the Eurovision Song Contest (ESC). This study explores the impact of musical features on the popularity of a song from the viewpoint of musical emotions. First, song popularity is approached using the emotional valence and arousal framework for low-level musical features. Then, high-level musical features are interpreted based on the concept of musical complexity. The data consists of 168 songs from four editions of the ESC from 2016 to 2019. Results indicate that while comprehensive predictions models of popularity based on emotions can be improved, the loudness and speechiness of a song can effectively predict its success at the ESC.
Introduction
“Step one: Believe in it And sing it all day long Step two: just roll with it That’s how you write a song” - Rybak, 2018, 1:25
These are the lyrics performed by the Norwegian Alexander Rybak “That’s How You Write A Song” at the Eurovision Song Contest (ESC) in 2018. The artist seduced the public in 2009 with his winning song Fairytale and returned in 2018. The truthfulness of the lyrics will not be discussed and are let to the reader’s discernment. This study emphasises the last sentence of the lyrics above, how to write a song, and particularly how to write a popular song. While believing in it and rolling with it might help, this paper focuses instead on the impact of musical characteristics in the public’s preference for a song.
The Eurovision1 is an international song competition taking place every year since 19562.
The contest consists mostly of European countries, each of them submitting a song that represents the nation in a live performance. The winner is determined through a combined tally of both a televote and a jury. Although there is no prize money for the winner, the country represented by the artist gets to host the next edition of the contest. Every year, betting websites give the opportunity to gamble on the scores and win money, and competitions are organised rewarding the model that is the most accurate in forecasting the winner (e.g. "Forecast Eurovision Voting | Kaggle", 2020).
1 The term ‘Eurovision’ is used to refer to the Eurovision Song Contest throughout this paper. 2 Barring the 2020 edition on the grounds of the COVID-19 pandemic.
There is evidence that the popularity of the songs can be predicted in part by politics via the so-called ‘political blocs’ (Yair, 1995; Gatherer, 2007). Certain countries vote for others with whom they have a political affiliation, thus giving rise to ‘alliances’ of countries voting for each other and ignoring the rest. A political approach to the Eurovision clearly sets aside the musical aspect of the contest. Or, in the words of one of the authors, “The winning song has no special traits: no superior harmonies, tunes or orchestration. In fact, the appreciation of music can have no objective rules, since songs reflect national taste, native rhythm and primordial meanings.” (Yair, 1995, p. 149). A second set of related writings concerns the cultural aspect of Eurovision (Yair, 1995, Gatherer, 2007; Ginsburgh & Noury, 2008). Similar to the political approach, a cultural view demonstrates the existence of voting blocs within the jury and “vote trading is rather the consequence of cultural factors” (Ginsburgh & Noury, 2008, p. 50). However, one point of critique is that this reasoning only applies to the judges while reforms of the voting system in 2016 strengthened the effect of the popular vote ("Rules", 2020). Recent studies have focused on the newly accrued importance of the public, for instance seeking to predict the success of a song based on how Twitter fans tweet about it (Demergis, 2019, October). Despite numerous studies on success at the Eurovision, the musical characteristics of the songs have been neglected as a predictor of popularity.
The ability to forecast the success of a song with precision is coveted by the music industry. Being able to tell which characteristics make a song popular is crucial for enterprises living off popular music such as record labels, radio stations and artists. A new body of literature exploring the impact of a song’s characteristics on its commercial success has emerged in the last two decades thanks to progress in data mining and machine learning. This body is commonly referred to as Hit Song Science (Dhanaraj & Logan, 2005, September; Ni et al., 2011, December; Pachet & Sony, 2012; Herremans et al., 2014; Pham et al., 2016; Middlebrook & Sheik, 2019) although the term is controversial in its definition (e.g. Ni et al., 2011,
December; Pachet & Sony, 2012. Hit Song Science studies are built on the assumption that popular songs share common musical features that can be measured, compared and analysed (Pachet & Sony, 2012). For instance, scholars have tried to predict the success of a song in the Billboard based on its musical characteristics (Parry, 2004; Herremans et al, 2014; Middlebrook & Sheik, 2019). There are various methods to obtain and analyse the musical features of a song, let it be a software (Parry, 2004) academic methods (Dhanaraj & Logan, 2005) or even Spotify data (Middlebrook & Sheik, 2019). Due to the variety of data it provides and the ease of access, the present study employs the tools available on Spotify for Developers (Home | Spotify for Developers, 2020).
Despite the ample number of studies conducted on the influence of musical characteristics on popularity, most of them focus on the Billboard ranking as a measure for success (e.g. Parry, 2004; Herremans et al, 2014; Middlebrook & Sheik, 2019). However, other external factors such as marketing techniques, ranking (Cibils, Meza & Ramel, 2015) or artist familiarity (Pham, Kyauk & Park, 2016) can be much more influential in predicting success than musical characteristics. The choice of the Eurovision is motivated by the contest showcasing mostly little-known artists with limited time for exposure before the contest, thereby making it less likely that success is determined by fame or promotion efforts. If proven conclusive, the results of this study will contribute to the body of research about musical characteristics and popularity while providing further evidence on which features matter and to what extent.
Finally, this paper aims to examine the impact of musical emotions and complexity on the popularity of a song at the Eurovision Song Contest. To do so, the first objective is to identify which features are most likely to impact popularity based on current literature, before analysing how they relate to musical emotions and preference.
Theoretical Background
Most of us should be familiar with listening to upbeat music that gets us excited, a melancholic song that makes us even sadder during a breakup or a quiet piece that puts us to sleep. Why do we like certain music and why does it make us feel a certain way? One could argue that our musical taste and how we react to different songs is very personal and hardly measurable in an objective manner. Others could add that although subjective, our preference for certain songs can be explained by its characteristics. This study holds the latter stance.
Musical Emotions
The field of psychoacoustics seeks to understand human responses to sound from a psychological standpoint. As human beings, we have a propensity to react to sound. Frequencies, notes or tones carry a meaning that can be associated with certain feelings when processed by our brain. As music is a mix of sounds, the listener may experience a complex reaction to it, and so music makes us feel emotions. This reasoning is the very premise of the concept of musical emotions, which aims at explaining how music stirs emotions (Gabrielsson, & Lindström, 2010). However, how this process occurs and to what extent is subject of debate. There are three principal theoretical currents focused on the interaction between music and humans (Coutinho & Cangelosi, 2011). The first approach states that music does not stir emotions in humans, as the emotion would fulfil no goal from a natural-survival standpoint. The reasoning applies despite music eventually carrying an emotional meaning. A second point of view defends the ability of music to generate emotions in its listeners but that these emotions are restrained to a set of emotional states and not basic emotions playing a survival role. A third perspective suggests that music can cause emotions to arise and that they are to be considered as ‘real’ emotions inasmuch as they trigger the same emotional mechanisms. The argumentation of the present paper follows the latter argumentation (Countinho & Cangelosi, 2011).
Building on the idea that music induces emotions, scholars have found evidence of different psychoacoustic features having an impact on emotions (Bruner, 1990; Streich, 2006; Gabrielsson, & Lindström, 2010; Lee & Lee, 2018). Research on the matter has made extensive use of the two-dimensional valence-arousal model (e.g., Coutinho & Cangelosi, 2011), where valence corresponds to the range of emotions perceived by the listener (e.g., happy vs. sad) and arousal is characterised by the intensity and energy of such emotions (Loui et al., 2013). In investigating which features have a prominent role on the inducement of emotion and their intensity, Coutinho and Cangelosi (2011) distinguished two categories of musical features. First low-level musical features or the basic characteristics of music (e.g., tempo, loudness) that are perceived similarly across individuals and are objectively measurable. Secondly, high-level features are subjective to the listener and shaped by the listener’s musical culture (e.g., modes). Table 1 shows a list of low-level musical features and their observed impact on musical emotions.
Table 1
Low-level musical features and musical emotions
Feature Levels Emotional expression
Tempo Fast Happy/Sparkling (Ga10; Co11), high emotional arousal (Co11) Slow Sad/Solemn (Br90; Ga10; Ca11), low emotional arousal (Co11) Loudness Overall Correlates with emotions arousal (Le18)
Loud Happy (Br90), High emotional arousal (Co11) Soft Sad (Br90), Low emotional arousal (Co11) Pitch High Happy (Br90), High emotional arousal (Co11)
Low Sad (Br90), Low emotional arousal (Co11) Voice Presence Higher emotional arousal (Lo13)
Absence Lower emotional arousal (Lo13)
Note. Authors’ names are shortened to the first two letters of the author and the two last digits of the publication year. Br90 = Bruner, 1990; Co11 = Countinho & Cangelosi, 2011; Le18 = Lee & Lee, 2018; Lo13 = Loui, Bachorik, Li & Schlaug, 2013).
Knowing that music stirs emotions in the listener, researchers have investigated the relationship between the valence of emotion and the preference for music. In other words, which type of emotions are more likely to translate into musical preferences. The hedonic consumption model explores the emotional impulse behind product use and preference as being bound to the listener’s experience or what is referred to as the ‘experiential view’ (Holbrook & Hirschman, 1982, p. 132 in Steininger & Gatzemeier, 2013). According to the model, emotion is one of the primary drivers of musical preference, and the effect is even stronger for pleasing emotions (Lacher, 1989). Emotionally pleasing emotions translate into positive emotions like joy or happiness, whereas unpleasant emotions are reflected by negative ones such as sadness or anger. In this case, musical preference could translate into the intention to vote for a song and consequently a song’s popularity. Therefore, the following hypothesis is formulated:
H1a: The more emotionally pleasing the music, the greater the popularity of the music.
In addition to emotional valence, the intensity of emotions – or arousal – also influences musical preference. Higher levels of arousal have been linked with increased preference for a song (Schäfer & Sedlmeier, 2010; Coutinho & Cangelosi, 2011; Lee & Lee, 2018). Thus, the following assumption is made:
H1b: Music with higher emotional arousal is more likely to be popular than music with lower emotional arousal.
Musical Complexity
When combined or modified, low-level musical features can form high-level musical measurement. The objective is to gain additional insights that single aspects would not have provided. One of the most noticeable uses of high-level musical features in the literature is referred to as musical complexity. In the words of Streich “the term complexity is complex” (Streich, 2006, p.13). Indeed, there is no unified answer as to what is meant exactly by musical
complexity. A mathematical approach of the concept defines it as “the measure of uncertainty in a musical situation” (Parry, 2004). In other words, if a rhythm or melody is very simple and repetitive, there is no uncertainty and thus no complexity. Complexity can be based on a single musical characteristic (e.g., Madsen & Widmer, 2006) or on the combination of many (see Lee & Lee, 2018). Diversity lies in the way in which musical complexity is calculated, and what musical features are included and put together. A few scholars have approached the notion of complexity from a diversity standpoint, where greater diversity would be synonym with increased complexity. This is the case of the studies carried out by Madsen & Widmer (2006) and Lee & Lee (2018) in which the authors make use of the Shannon entropy, developed by the mathematician of the same name (Shannon, 1948). The entropy is used to calculate the uncertainty of predicting the musical characteristics. In other words, the more uncertain the features of a song are to predict, the more likely they are to be complex. Researchers have found evidence that musical complexity is correlated with musical preference insofar as higher complexity is linked with increased preference for a song (Steck & Machotka 1975; Parry, 2004). Table 2 describes how different types of musical complexity relate to musical emotions: Table 2
High-level musical features and musical emotions
Feature Levels Emotional expression
Change in rhythm Large Low level of emotional arousal (St06) Moderate High level of emotional arousal (St06) Small Low level of emotional arousal (St06) Change in melody Large Happiness, pleasantness, surprise (Ga10)
Small Disgust, anger fear boredom (Ga10) Change in loudness Large Low emotional arousal (Le18)
Small High emotional arousal (Le18)
Note. Authors’ names are shortened to the first two letters of the author and the two last digits of the publication year (e.g., Ga10 = Gabrielsson & Lindström, 2010).
Studies have shown that musical complexity correlates with the popularity of a song (Madsen & Widmer, 2007, January; Gatherer, 2007; Gabrielsson & Lindström, 2010 Lee & Lee, 2018). The literature suggests however that the relationship between the two might not be linear but rather follows an inverted U-shape relationship (Streich, 2006; Coutinho & Cangelosi, 2011). Considering the latter finding, it seems plausible that songs with moderate levels of musical complexity are more popular than songs with low or high levels of complexity. Therefore, this study aims to examine whether there is an association between musical complexity and song popularity and if so, the nature of their relationship.
Methods Sample & Research Design
The dataset consisted of 168 songs that were qualified for the semi-finals and finals of the Eurovision Song Contest for the four editions between 2016 and 20193. Songs before 2016
we excluded because the voting system changed that year, giving more importance to the televote score thus rendering the comparison with previous years unreliable. Hence the editions prior to 2016 were not chosen in order to limit motivations other than musical and comparison unreliability in score computation. Although the jury vote has proven to be culturally and politically motivated rather than based on the sheer quality of the song (see e.g. Yair, 1995), it was included in the analysis as it provides a different perspective on song popularity. The dataset of the scores for each country, song and edition was obtained on the data science website Kaggle ("Eurovision Song Contest scores 1975-2019", 2020). The name of the song, artist, album as well as country, language, gender and group composition were manually coded from the Eurovision’s official website. The Spotify ID – a code specific to each song on Spotify –
was subsequently retrieved via the Spotify Web API ("Web API, Spotify for Developers", 2020). The same tool was used to extract the audio features for each song from Spotify for both the low-level and high-level acoustic features (see Appendices). The analyses were conducted with version 25 of the statistical software SPSS. The tests were realised using independents sample T-tests, bivariate correlations, linear regression as well as the curve estimation. The dataset is accessible at the following link: https://tinyurl.com/ya87kdlh. None of the 168 songs that competed in the editions from 2016 to 2019 was missing. Table 3 provides an account of the number of countries participating per edition. According to the rules of Eurovision, a total of six countries is automatically qualified to the final for historical reasons: the host country, as well as France, Germany, Spain, Italy and the United Kingdom. The countries qualified to the final are announced at the end of the semi-final, but not their points. The scores for each song are revealed to the viewers at the end of the final. The songs and scores from the semi-finals and final were included in the dataset. On average, the songs were slightly over 3 minutes long (M = 182.26, SD = 88.08) and received 187 points in the finals if qualified (M = 186.62, SD = 147.03). Amar Pelos Dois by Salvador Sobral from Portugal in 2017 is the song with the most points ever earned in the final, with 758 points. On the other hand, Spirit of the Night by Monetta & Wilson from San Marino received only 1 point in the semi-final in 2017.
Table 3
Number of Countries Participating per Edition
Edition Semi-final Final Total
2016 36 26 42
2017 36 26 42
2018 37 26 43
2019 35 26 41
Total 144 104 168
Measures
Emotional valence
The concept of emotional valence refers to the range of emotions perceived by the listener (Loui et al., 2013). Emotions are characterised on a linear scale from negative (e.g. anger, sadness) to positive (e.g. joy, happiness). Emotional valence was constructed based on tempo, loudness, speechiness, pitch and valence. Emotional valence was computed based on the average value of the five variables. The resulting scale ranged from 0 to 1, where a value of 0 represents negative emotional valence and 1 represents positive emotional valence. Each of the five items was computed as described below.
Tempo is a measure of the overall estimated tempo of a track in beats per minute (BPM). The scale for tempo was adjusted to a 0 to 1 scale by coding what is considered as a very slow tempo (20BPM) as 0, and very fast tempo (220BPM) as 1. All the values fit into this interval (M = 0.50, SD = 0.14).
Loudness is a measure of the overall loudness of a track in decibels (dB). Loudness as provided by Spotify is measured on a negative scale, where 0 dB corresponds to the original volume of the song and any value below that represent a decrease in the loudness of the signal on a logarithmic scale. The variable was recoded into a 0 to 1 scale, where a softer song (-20dB) was recoded as 0 and louder song (0dB) as 1 (M = 0.69, SD = 0.11).
Speechiness detects the presence of spoken words in a track. The more prominent the lyrics are, the closer to 1.0 the attribute value (M = 0.05, SD = 0.03). Higher values refer to tracks mainly made of spoken words and lower values represent music and non-speech tracks.
Pitch: the overall key of a song, for instance, C, C# to B. Since keys are ordinal, all values were recoded from a 0 (lower pitch) to 1 (higher pitch) scale (M = 0.49, SD = 0.34).
Valence was measured from 0 to 1, describing the musical positiveness conveyed by a song. Songs with higher values have a more positive valence (e.g. happy, cheerful, euphoric),
while tracks with lower values have a lower valence and sound more negative (e.g. sad, depressed, angry) (M = 0.41, SD = 0.19).
A factor analysis was conducted on the five variables composing emotional valence using Principal Axis Factoring and direct Oblimin (oblique rotation). The KMO measure for sampling adequacy was above the minimum criterion of 0.5 (KMO = .60), and Bartlett’s test of sphericity shows the correlations between variables to be significantly different from zero, χ2 (168) = 56.91, p < .001. Based on Kaiser’s criterion and the scree plot, one factor was retained with an Eigenvalue of 1.70 explaining 34.0% of the variance in the individual items. All factors were retained based on their factor loadings at the exception of tempo which had a loading below .10. A reliability analysis with valence, loudness, pitch and speechiness revealed that the scale was not reliable and does not improve in terms of reliability if an item is removed (α = .31). Therefore, all items will be treated separately in the analysis.
Emotional arousal
The concept of emotional arousal represents the intensity and energy of emotions perceived by the listener and was measured using the variable energy provided by Spotify. Energy was measured on a scale from 0.0 to 1.0, representing a perceptual measure of intensity and activity, with 0 indicating low energy and 1 indicating high energy (M = 0.66, SD = 0.17). Musical complexity
The concept of musical complexity refers to the uncertainty of a given note, pitch, timbre, etc, to occur. Uncertainty is measured via the Shannon diversity index. Songs with a low index value have little diversity, meaning that their musical characteristics are very predictable and subsequently the song has low musical complexity. A high index value indicates higher musical complexity. For instance, a song with each section in C# is not diverse and therefore has very low musical complexity. For this study, musical complexity is measured by the diversity of pitch classes (e.g. C#, D) of different time intervals called sections. There are 12 pitch classes
from 0 to 11. The more musically complex a song is, the higher the value, with the lowest being 0 (M = 1.02, SD = 0.32). A lower score stands for low musical complexity and a higher score for high complexity. A section is defined by large variations in rhythm or timbre, e.g. chorus, verse, bridge, guitar solo, etc. Each section is given a unique pitch class.
Popularity
The popularity of a song is defined by the total amount of points obtained by a song. There are two ways a song can receive points. The first is from televoting, where viewers can vote by SMS or via the ESC app for their favourite song. It is not possible to vote for your own country. For each country, the song that has received the most votes is ranked first and gets 12 points, the second most popular 10 points, the third most popular 8 points, the fourth most popular 7 points and so on. As an example, Soldi by Mahmood of Italy has obtained the most votes in Belgium, so Italy received 12 points from the televote of Belgium. The second way to obtain points if from the national jury. Each represented country brings its own jury to the contest and votes for the contestants of other countries. The point system is the same as for the televote, 12 then 10, 8, 7, etc. Both the televote and jury vote are then added to form the total score of a song in the semi-finals and final. Table 4 shows the six indicators of popularity. Table 4
Descriptive Statistics of the Measures of Popularity
n M SD Semi-final Televote 144 56.73 52.82 Jury 144 57.05 49.54 Total 144 113.77 91.72 Final Televote 104 57.76 87.01 Jury 104 57.76 75.18 Total 104 115.52 146.95
Covariates
A series of covariates were selected and added to the analysis based on the literature. They concern either the language of the song or the characteristics of the performers. The variables are the following:
Language: the primary language in which the track is sung. Most of the songs are sung in English (79.8%) followed by Italian (3.0%) and French and Croatian (2.4%).
Secondary Language: the secondary language in which the track is sung. Most of the songs were sung only in one language (94.0%) and the rest was bilingual.
Gender: the gender of the performer(s) of the song. In the sample, 44.0% of the performances showcased only males (n = 74), 42.9% female (n = 72) and the rest (13.1%) were mixed (n = 22).
Number of Performers: the number of performers in the song. Most of the songs had a single performer (76.8%), and a minority (23.2%) had multiple performers on stage.
Results Emotional Valence
The results in Table 5 show that all items of emotional valence were significantly related to each other except for tempo. Loudness, speechiness, pitch and valence positively correlated with each other. In other words, an increase in one of the four items is likely to be associated with an increase in the other three. On the other hand, tempo did not correlate significantly with the other items. This means that valence, loudness, speechiness and pitch are likely to measure a similar construct while tempo measures a second one. While the factor analysis was not successful, the correlation shows that a possible scale for emotional valence could include some of the items chosen for the analysis.
Table 5
Descriptive Statistics and Correlations for Variables of Emotional Valence
Variable n M SD 1 2 3 4 5 1. Valence 168 0.41 0.19 - 2. Tempo 168 119.73 27.01 -.04 - 3. Loudness 168 -6.29 2.24 .32* .03 - 4. Speechiness 168 0.05 0.03 .34* .13 .22* - 5. Pitch 168 5.39 3.73 .20** .13 .03 .23** - Note. *p < .001., ** p < .05
In order to determine the effect of loudness, speechiness, pitch and valence on song popularity, a multiple regression analysis was conducted for each of the outcome variables of song popularity. The results in Table 6 show that of the six measures of song popularity, two showed significant results. First, the model predicting the score of the televote for the final was statistically significant, F (4,163) = 3.27, p = .013.Together, the predictors explain 7.4% of the variations in the score of the televote in the final, which is relatively low, R² = .07. Loudness negatively predicts the score, b = -161.19. The partial effect of the variable is statistically significant, t = -2.59, p = .011, 95% CI [-284.33; -38.04] and the effect is moderate, b* = -.21. Speechiness positively predicts the score, b = 554.28. The effect of speechiness on televote for the final is statistically significant, t = 2.23, p = .027, 95% CI [63.33;1045.23] and the effect is moderate, b* = .18. The second model, predicting the total score for the final, was statistically significant, F (4,163) = 2.71, p = .032. Overall, the predictors account for 6.2% of the variations in the score in the final, R² = .06. Loudness negatively predicts the score, b = -264.97. The partial effect is statistically significant, t = -2.50, p = .013, 95% CI [-474.29; -55.65]. The other variables were not found to have a significant effect on the outcome variable. The analysis showed that emotional valence cannot be reliably measured as a unified construct using the
chosen items. However, the items do influence one another as demonstrated by the correlation analyses. The hypothesis H1a was rejected.
Table 6
Model Comparison of Popularity with Predictors of Emotional Valence*
Outcome Variable n F p R² Semi-finals Jury 144 1.86 .105 0.02 Televote 144 0.56 .728 0.05 Total 144 1.28 .275 0.04 Final Jury 104 1.16 .333 0.03 Televote 104 2.60 .027 0.07 Total 104 2.71 .032 0.06
*Note. As predictors, valence, pitch, loudness, speechiness and tempo were added into the models.
Furthermore, the model with televote and total score of the semi-final provide empirical support for the influence of speechiness on popularity in that more salient lyrics positively affect the score. In sum, the assumption that songs with positive emotional valence are more popular than songs with negative emotional valence is not confirmed. The selected musical features have been found to be unable to homogeneously measure emotional valence. Moreover, the analysis showed that the findings and the theory do not necessarily align in terms of how musical features translate into emotional valence. Hence the hypothesis that songs with positive emotional valence are more popular than songs with negative emotional valence is rejected. Although the results are inconclusive, the bar graph in Figure 1 demonstrates that the model with tempo, speechiness, loudness, pitch and valence as predictors has higher predictive power for the televote than for the jury vote. As expected, a clear distinction is observed between the predictability of the jury score and the televote score, with the Televote score being better predicted by emotional valence than the jury score.
Figure 1
Comparison of the Emotional Valence Model
Emotional Arousal
To test the hypothesis that higher levels of emotional arousal were linked to higher song popularity, a correlation analysis was conducted. The results indicate that there was no significant correlation between energy and the variables of popularity as demonstrated in Table 7. The Pearson correlation values seem to indicate very weak associations and multiple directions possible, as the values ranged from .09 to -.07. In short, the assumption that songs with higher emotional arousal are more popular than songs with lower emotional arousal is rejected, as there is no evidence of an association between energy and the variables for popularity. Since there was no correlation established between emotional arousal and song popularity, the direction of the relationship was not further investigated. The hypothesis H1b is rejected.
Table 7
Correlation for Emotional Arousal and Song Popularity
Variable n M SD Energy Energy 168 0.66 0.17 - Televote Semi-final 144 56.73 52.81 .09 Jury Semi-final 144 57.05 49.54 -.03 Total Semi-final 144 113.77 91.71 .03 Televote Final 104 57.76 87.01 -.01 Jury Final 104 57.76 75.18 -.13 Total Final 104 115.52 146.95 -.07
Note. The results of the correlation analysis were not statistically significant.
Musical Complexity
Studies on musical complexity and song popularity have provided evidence of a effect of musical complexity on song popularity. A linear regression was run to test this proposition. The outcome of the regression analysis in Table 8 indicate that there is no statistically significant effect of melodic complexity on any of the variables for song popularity.
Although the linear correlation did not prove conclusive, the literature on musical complexity alluded to an inverted U-shape relationship with song popularity. To test that assumption, a quadratic curve estimation was conducted in order to find whether there was a significant non-linear correlation between musical complexity and the variables of song popularity. The analysis reveals that the relationship is statistically not significant, F (2,165) = 1.08, p = .344. Besides, the fit of the curve was weak with R = .01. The results do not provide empirical evidence to the idea that extreme levels of musical complexity have a negative impact on song popularity compared to moderate levels of complexity having a positive impact.
Table 8
Regression Analysis for Musical Complexity and Song Popularity
Variable n M SD Melodic Complexity p
R² Melodic Complexity 168 1.02 0.32 - - Televote Semi-final 144 56.73 52.81 .00 .472 Jury Semi-final 144 57.05 49.54 .01 .334 Total Semi-final 144 113.77 91.71 .01 .349 Televote Final 104 57.76 87.01 .01 .177 Jury Final 104 57.76 75.18 .00 .478 Total Final 104 115.52 146.95 .001 .245 Covariates
An independent samples T-test was conducted with the variables of song popularity as outcome and number of performers as a grouping variable. There is a statistically significant difference in score between solo performances and group performances for the jury vote in both the semi-finals and final. For the semi-finals, the test revealed that scores were significantly higher for solo performances (M = 60.94, SD = 50.79) than group performances (M = 44.18, SD = 43.29). The Leven’s test for equality of variance was statistically significant meaning that the equality of variances was not assumed (F = 10.30, p = .009), t (72.49) = 2.03, p = .046, d = 0.36, 95% CI [0.32, 33.20]. The same results were found for the jury vote in the final, where scores were significantly higher for solo performances (M = 63.88, SD = 81.28) than for group performances (M = 37.51, SD = 45.11), t = (116.01) = 2.59, p = .011, d = 0.40, 95% CI [6.23, 46.51]. The effect sizes are small to medium and reveal that on average the jury is more likely to award more points to songs with a solo performer than with a group performance.
A One-way ANOVA was conducted to compare the effect of gender on song popularity in the female, male and mixed conditions. The analysis demonstrates that there was a significant difference of gender on the televote score in the final, F (2,165) = 3.35, p = .037, as well as on the jury score in the final F (2,165) = 4.06, p = .019, and the final score F (2,165) = 4.24, p = .016. Post-Hoc comparison using the Bonferroni test indicated that the mean score for males (M = 77.08, SD = 98.23) was higher than for females (M = 42.54, SD = 73.67) for the televote score in the final. In the case of the jury score in the final, the mean score for male (M = 73.93, SD = 86.53) and mixed (M = 26.68, SD = 37.40) differed significantly. Finally, the mean score for male (M = 151.01, SD = 164.47) and female (M = 93.18, SD = 131.87) differed significantly for the total score in the final.
Two additional independent samples T-test were conducted. The first one tested the effect of language on song popularity in the English and non-English conditions. The second tested the effect of bilingualism on song popularity in the bilingual and non-bilingual conditions. Both analyses found no statistically significant mean differences between the groups. In sum, the analysis revealed that the gender and of the performers and their number ultimately matter for the final score. The results indicate a preference for male performers over female and mixed a solo performance as well over group performances. The language in which the song was performed was not found to influence the final score.
Discussion
The general aim of the study was to explore the effect of musical emotions and complexity on the popularity of a song at the Eurovision Song Contest. The results suggest that musical emotions do not have a clear-cut impact on song popularity. The first analysis of emotional valence demonstrated that the salience of the lyrics in a song had a positive impact, and loudness had a negative impact on the song’s popularity. However, individual music characteristics could
not be combined on a single scale meaningfully according to their valence as indicated in the literature. Second, the emotional arousal of a song was not found to have a significant impact on popularity. A correlation analysis did not reveal any association between the two variables. Lastly, the results did not support any association between musical complexity and song popularity. The analysis of the covariates showed that the gender and the number of performers had an impact on the final score, with solo male performers obtaining higher scores on average than any other groups. On the other hand, the languages in which the song was performed did not influence the final score. Altogether, the chosen musical characteristics and emotions were not found to have a consequential predictive power as for why we prefer certain songs. To conclude, the negative results of the effect of musical emotions on song popularity should not be interpreted as the demonstration that popularity cannot be predicted from an analysis of musical features. Rather, the selected factors and framework do not provide enough insight into the subjective judgment behind song preferences.
Based on the measure of speechiness obtained via Spotify, we observed that the salience of the lyrics in a song was leading to a higher total score. The effect was shared across both the televote and the jury vote, meaning that the influence is likely to be generalisable among both bodies of voting. The fact that lyrics played a significant role in the public’s preference for a song can be explained by the importance of lyrics at the Eurovision. The public is more likely to prefer songs with recognisable or catchy lyrics. For instance, Amar Pelos Dois by Salvador Sobral won the contest in 2017. The song is a sad and melancholic love song with nothing but piano and strings in the background. The singer managed to obtain an important number of points, eventually leading him to the top spot. The results also showed that the loudness of a song was negatively associated with popularity, which is contrary to what was stipulated in the literature. Studies have provided evidence that loudness is positively linked to popularity. Reflecting on the rationale behind such results, softer songs may be more popular because they
are less overwhelming and allow the listener to identify clearly the lyrics and other catchy features of a song. A good example – once again – is Amar Pelos Dois which is the softest song of all, or Arcade by Duncan Laurence who won in 2019 and is the fourth softest song of the dataset. Overall, the fact loudness and speechiness had an impact on the song’s popularity is double-edged. It shows that musical features are associated with popularity; and exposes a divergence between the theory and what has been observed in this paper. Both of which call for further research on the matter.
When controlling for gender and number of performers, a series of group comparisons revealed that both the jury and the public had observable inclinations concerning who was performing. Male performers received on average 58 points more than their counterpart in the final. A similar gap occurs between solo performers and group performers, in favour of solo artists. The special affinity for solo performances is reflected in the dataset, where all winning sounds are solo performances. Group performances may make it harder for the audience to focus on the song and lyrics in themselves. Such results are echoing the preferences for soft songs with dominant lyrics.
Limitations
The present study was an attempt to interpret the popularity of a song not merely from a technical standpoint but also a psychological perspective. It was made clear by the results that the paper possesses some shortcomings. To begin with, the theoretical framework about music and emotions can be criticised on different grounds. Firstly, assumptions were made that voters at the Eurovision tend to vote for songs that make them feel positive emotions rather than negative ones, which is not necessarily the case. Secondly, evidence was provided that voting can be motivated by other reasons than the aesthetics of a song, such as politics and culture. Lastly, the framework does not account for influences other than the acoustic characteristics of a song. The Eurovision is a highly visual contest and surely the light show, dance, costumes,
etc (i.e., the performance) had an impact on the scores. Additionally, some methodical limitations are to be underlined. The Spotify API gives the opportunity to access audio data relatively easily but does not provide extensive details regarding the manner in which the variables are calculated, raising concerns in terms of transparency for variables. Musical complexity was interpreted in the simplest form found in the literature, but other researchers have done a more comprehensive and thorough analysis of the subject (e.g. Streich, 2006). The sample consist of 168 songs for reasons explained previously. While the size is satisfactory it would benefit to be bigger, would related studies explore moderating and mediating factors.
Future studies include building a grounded theoretical framework on the relationship between various musical features and their translation into emotions. Developing a reliable scale for emotional valence and arousal based on these features would be beneficial to gain a deeper knowledge of how the features of a song influence our preference through emotions. With the rise of hit song science and machine learning, researchers could look for meaningful variables predicting the popularity of a song for the contest, potentially adding the data from previous and future years.
Conclusion
This study proposed methods for measuring emotions and complexity in music as a mean to predict the popularity of a song at the Eurovision. The results did not provide support for a comprehensive prediction model of popularity based on emotions and complexity. Nevertheless, the loudness and speechiness of a song were proven effective in predicting its success. There is room for improvement in models predicting popularity but the fact that musical preferences may be explained in part by characteristics inherent to the music calls for further investigation. There is still no answer on how to write a popular song at the Eurovision. One thing is for certain, musical characteristics matter.
References
Bruner, G. C. (1990). Music, mood, and marketing. Journal of Marketing, 54(4), 94-104. Cibils, C., Meza, Z., & Ramel, G. (2015). [Unpublished Manuscript]. Predicting a song's path
through the Billboard Hot 100'. Stanford University
Coutinho, E., & Cangelosi, A. (2011). Musical emotions: predicting second-by-second
subjective feelings of emotion from low-level psychoacoustic features and physiological measurements. Emotion, 11(4), 921.
Demergis, D. (2019, October). Predicting Eurovision Song Contest results by interpreting the tweets of Eurovision fans. In 2019 Sixth International Conference on Social Networks Analysis, Management and Security (SNAMS) (pp. 521-528). IEEE.
Dhanaraj, R., & Logan, B. (2005, September). Automatic prediction of hit songs. In ISMIR (pp. 488-491).
Eurovision Song Contest scores 1975-2019. (2020). Retrieved 1 June 2020, from https://www.kaggle.com/datagraver/eurovision-song-contest-scores-19752019. Forecast Eurovision voting | Kaggle. (2020). Retrieved 1 June 2020, from
https://www.kaggle.com/c/Eurovision2010.
Gabrielsson, A., & Lindström, E. (2010). The role of structure in the musical expression of emotions. Handbook of music and emotion: Theory, research, applications, 367400. Gatherer, D. (2007). Voting in Eurovision: shared tastes or cultural epidemic?. Empirical Text
and Culture Research, 3, 72-83.
Ginsburgh, V., & Noury, A. G. (2008). The Eurovision song contest. Is voting political or cultural?. European Journal of Political Economy, 24(1), 41-52.
Herremans, D., Martens, D., & Sörensen, K. (2014). Dance hit song prediction. Journal of New Music Research, 43(3), 291-302.
Home | Spotify for Developers. (2020). Retrieved 1 June 2020, from https://developer.spotify.com/.
Lacher, K. T. (1989). Hedonic consumption: Music as a product. Advances in consumer research, 16(1), 367-373.
Lee, J., & Lee, J. S. (2018). Music popularity: Metrics, characteristics, and audio-based prediction. IEEE Transactions on Multimedia, 20(11), 3173-3182.
Loui, P., Bachorik, J., Li, H. C., & Schlaug, G. (2013). Effects of voice on emotional arousal. Frontiers in psychology, 4, 675.
Madsen, S. T., & Widmer, G. (2006). Music complexity measures predicting the listening experience. Entropy, 1(4), 1.
Madsen, S. T., & Widmer, G. (2007, January). A complexity-based approach to melody track identification in midi files. In Proceedings of the International Workshop on Artificial Intelligence and Music.
Middlebrook, K., & Sheik, K. (2019). Song Hit Prediction: Predicting Billboard Hits Using Spotify Data. arXiv preprint arXiv:1908.08609.
Ni, Y., Santos-Rodriguez, R., Mcvicar, M., & De Bie, T. (2011, December). Hit song science once again a science. In 4th International Workshop on Machine Learning and Music: Learning from Musical Structure, Sierra Nevada, Spain.
Pachet, F., & Sony, C. S. L. (2012). Hit song science. Music data mining, 305-326. Parry, R. M. (2004). Musical complexity and top 40 chart performance. Georgia Institute of Technology.
Parry, R. M. (2004). Musical complexity and top 40 chart performance. Georgia Institute of Technology.
Schäfer, T., & Sedlmeier, P. (2010). What makes us like music? Determinants of music preference. Psychology of Aesthetics, Creativity, and the Arts, 4(4), 223.
Rules. (2020). Retrieved 1 June 2020, from https://eurovision.tv/about/rules/. Rybak, A. (2018). That's How You Write A Song. Retrieved 15 June 2020, from
https://www.youtube.com/watch?time_continue=85&v=SixPA1jhynE&feature=emb_tit le.
Shannon, C. E. (1948). A mathematical theory of communication. Bell system technical journal, 27(3), 379-423.
Steck, L., & Machotka, P. (1975). Preference for musical complexity: Effects of context. Journal of Experimental Psychology: Human Perception and Performance, 1(2), 170. Steininger, D. M., & Gatzemeier, S. (2013). Using the wisdom of the crowd to predict popular
music chart success. In ECIS (p. 215).
Streich, S. (2006). Music complexity: a multi-faceted description of audio content. Barcelona, Spain: Universitat Pompeu Fabra.
Yair, G. (1995). ‘Unite Unite Europe’ The political and cultural structures of Europe as reflected in the Eurovision Song Contest. Social Networks, 17(2), 147-161. Web API | Spotify for Developers. (2020). Retrieved 1 June 2020, from
Appendices
