Open-Source Intelligence and Music Information Retrieval for Geographic Attribution of Musical Affect and the Ecological Limits of Population Inference Mohammadreza Rashidi
arXiv:2607.12517v1 [cs.CR] 14 Jul 2026
Department of Computer Science AI and Media Analysis Lab Berlin, Germany [email protected] Abstract—A common intuition holds that a region’s music mirrors the temperament of its people, so that melancholic melodies mark melancholic populations. We test the measurable half of that intuition and reject the inferential half. Using the Essen Folksong Collection, a corpus of thousands of notated folk melodies, we extract real melodic and affectrelated features from 2393 deduplicated melodies spanning 16 countries and 7 geographic regions, with the analysis performed on symbolic scores rather than audio. The mode of each melody is computed with a key-finding algorithm rather than read from the file, because the collection’s own documentation warns its major and minor labels are unreliable. Cross-country differences in melodic structure are large and highly significant. All 8 tested features differ across countries at p < 0.001, with the leap-related features reaching p < 10−90 , and China carries a distinctive wide-leap, highactivity signature (arousal composite +1.24 standard deviations, mean absolute interval 2.77 semitones against Germany’s 2.17). We then test the inferential half. We correlate the regional musical-affect measures with two published, validated national indices, the World Happiness Report ladder score and the Hofstede individualism index. None of the 6 correlations is significant (0 of 6). The geography of musical affect is real and measurable, but it does not predict how happy or how individualist a population is, and any claim that it does is an ecological fallacy. We release the full extraction and analysis pipeline, and a fail-closed checker re-derives every number in this paper from the data. Index Terms—computational musicology, cross-cultural music, music and emotion, folk song, ecological fallacy, symbolic music analysis
1. Introduction It is an old and appealing idea that the music of a place expresses the soul of its people. Tourists describe Portuguese fado as proof that the Portuguese are saudade-laden, and a listener who finds a folk tune plaintive is quick to read the same quality into the community that sang it. The idea is appealing because one half of it is true and measurable. Music does carry affect, and the affect of a region’s music
can be quantified. The other half, the inference from the music to the psychology of the people, is where the appeal becomes a trap. To move from “this region’s songs use more of a sad-sounding device” to “this region’s people are sad” is to commit the ecological fallacy [1], reasoning from a group-level aggregate to the individuals in it, and to trade measurement for stereotype. This paper separates the two halves and reports both honestly. We measure the first and reject the second. On the measurement side, we take the Essen Folksong Collection [2], a large corpus of folk melodies encoded as symbolic scores, and extract melodic and affect-related features from 2393 deduplicated melodies across 16 countries grouped into 7 geographic regions. On the inference side, we ask whether the resulting regional musical-affect profiles predict two published, validated national indices that do describe people, the World Happiness Report ladder score [3] and the Hofstede individualism index [4]. They do not. Our contributions are three. •
•
•
A reproducible, symbolic-score analysis of the affective structure of folk melodies by geography, showing that all 8 tested features differ across countries at p < 0.001 and isolating region-specific signatures such as China’s wide-leap, high-activity profile. A rigorous negative result. The regional musicalaffect measures do not significantly correlate with national happiness or individualism (0 of 6 correlations significant), which directly rebuts the folk inference from a region’s music to its people’s disposition. An explicit methodological stance against the ecological fallacy and the unreliable-mode and corpusimbalance pitfalls that make naive versions of this study produce confident but meaningless claims, together with a released pipeline and a fail-closed numeric checker.
2. Background and Related Work Computational and cross-cultural musicology. Symbolic music analysis has a mature toolchain. We use music21 [5] to parse the Humdrum **kern scores of
the Essen collection [2], which encodes thousands of folk melodies with geographic tags. Large-scale cross-cultural work has established that human music has both statistical universals and systematic regional variation [6], [7], which is the backdrop against which our regional differences should be read. We expect and find real variation, and the question is what may and may not be inferred from it. Music and emotion. That specific musical features carry affect is one of the most robust findings in music psychology. Mode and tempo are the two strongest cues to the happy-sad axis, with major and faster reading as happier and minor and slower as sadder [8], [9], and a broad review identifies a shared code of acoustic and structural cues to emotion across vocal expression and music [10]. Affect itself is commonly modelled in two dimensions, valence and arousal, following Russell’s circumplex [11], and the dimensional model fits music-emotion data well [12]. Our two composites are deliberately aligned with these dimensions. The wider task of predicting emotion from features, music emotion recognition, is a mature subfield with its own reviews and feature sets [13], [14], and our structural features overlap with the melodic features it uses. We adopt the established cue directions not to assign a definitive emotion to any melody, but to orient two transparent composite measures of valence and arousal. Expectation and probabilistic structure. A complementary tradition models melody as a probabilistic sequence, where the entropy of the note distribution and the predictability of the contour carry both stylistic and affective weight [15], [16]. Our pitch-class entropy feature is a direct instance of this view, and its separation of pentatonic from diatonic idioms is the kind of structural signature that tradition predicts. The ecological fallacy. Robinson’s classic result [1] showed that correlations computed on group aggregates can differ in sign and magnitude from the individual-level correlations they are taken to represent. Inferring a population’s temperament from the aggregate affect of its music is a textbook instance, and our negative result is, in effect, a demonstration that the aggregate-level correlation is not even present, let alone transferable to individuals.
3. Data Sources. We draw on four real, public symbolic corpora, chosen so that every melody analysed is a genuine transcription rather than a synthetic or paraphrased example. The Essen Folksong Collection [2] is a large body of folk melodies in the Humdrum **kern format, arranged by continent and then by country, and it supplies all the continental European material, the Chinese folk subset, and a small United States set. The Nottingham Music Database [17] supplies British folk melodies in MIDI, giving a second, independently encoded folk source outside the Essen collection. The ShourCorpus [18] supplies non-metric Iranian classical melodies of the Dastgah Shour in MIDI, and the SymbTr corpus [19] supplies Turkish makam melodies in MIDI. The four corpora are independently created and encoded,
which is a strength for the folk-versus-classical contrast and a caveat for any direct like-for-like comparison, and we treat it accordingly. The Essen Folksong Collection [2] is arranged by continent and then by country. We analyse the countries with enough melodies to estimate a distribution, listed in Table 1. Several properties of the corpus shape the analysis and are stated plainly rather than hidden. First, the collection is heavily weighted toward Germany and China, with many European countries represented by tens rather than hundreds of songs. To keep a cross-country comparison from being dominated by the two largest countries, we cap each country at a fixed random sample of 400 melodies drawn with a fixed seed, and we use all available melodies for the smaller countries. Second, the collection contains many variant renderings of the same song, which would violate the independence of observations. We deduplicate by title within each country before sampling. Third, the corpus is a collection of traditional melodies, not a sample of what people in these countries listen to today, and its European material is copyright of the Schaffrath estate and used here for research analysis only, with no melodies redistributed. Beyond continental Europe and China, the Essen collection also carries a small United States folk set, which we keep in the descriptive table for geographic breadth but flag as too few melodies for a stable distribution, so it enters the figures and the per-country profile but not the significance tests, the benchmark, or the index correlation. For a second, independently encoded folk source we add the British melodies of the Nottingham Music Database [17], capped by the same rule as the Essen countries, which gives Western Europe a source outside the Schaffrath collection. The Essen collection covers Europe and China but contains no Middle Eastern material. To include Middle Eastern traditions with real, public symbolic sources, we add two. The 29 pieces of the ShourCorpus of non-metric Iranian classical music [18], and a 400-piece sample of the SymbTr corpus of Turkish makam music [19]. These are deliberately cautious additions, and we flag their differences rather than blend them in silently. Both are 2 classical, largely nonmetric, and microtonal art-music traditions, and their MIDI encodings collapse quarter-tones and commas to the nearest semitone, so their interval statistics are approximations in a twelve-tone frame that the music does not use. We therefore include Iran and Turkey in the descriptive profile, the distance and principal-component maps, and the figures, but exclude them from the folk-only significance tests and the index correlation, which keep a single repertoire and encoding. For a Jewish and Israeli symbolic tradition we add the Beregovski corpus of klezmer, the Ashkenazi Jewish instrumental folk music transcribed to the same Humdrum **kern format as the Essen collection by the Mode-inKlezmer project [20], [21], which yields 94 melodies after parsing and the same deduplication and note-count filters. We label this row Israel because klezmer is central to Jewish and Israeli musical tradition, but we state plainly what it is
and is not. It is instrumental dance music rather than vocal folk song, it is Ashkenazi and Eastern European in origin rather than Mizrahi or modern Israeli popular music, and its characteristic modes such as freygish are neither major nor minor, so we treat its minor-mode fraction as an unreliable valence cue exactly as we do for the Chinese pentatonic subset. Because the encoding is diatonic and metric and matches the Essen material, we include Israel in the folk statistics, the benchmark, and the arousal index correlation, while flagging that its position may reflect the instrumental idiom and Ashkenazi origin as much as modern geography. Japan, India, and Pakistan, which we also sought so as to widen the Asian coverage, have no comparable public symbolic corpus that we could obtain under an open license, and we leave them out rather than fabricate data or scrape audio of unknown provenance.
4. Method Feature extraction. From each melody we compute seven structural features that do not depend on a tonalmode label and are therefore comparable across idioms. Mean pitch, pitch range, mean absolute melodic interval, leap ratio (the fraction of intervals larger than a whole tone), ascending ratio, note density (notes per unit score time), and pitch-class entropy. We additionally compute the mode (major or minor) with the Krumhansl-Schmuckler key-finding algorithm [22] as implemented in music21 [5]. We compute the mode rather than read it from the file because the collection’s documentation states that its own major and minor designations are unreliable. We return to a deeper problem with mode in the discussion. A forced major or minor label is not a meaningful valence cue for a melody that does not live in the major-minor system at all, such as a Chinese pentatonic tune. Affect composites. We form two composite measures from the country-level feature means, standardised across countries. The arousal composite averages the z-scores of note density, mean absolute interval, and pitch range, following the established association of activity, large intervals, and wide range with higher arousal [10]. The valence composite averages the z-scores of major-mode fraction, mean pitch, and ascending ratio, following the mode and register cues to the happy-sad axis [8], [9]. We report the raw features alongside the composites so that nothing in the analysis depends on the composite definitions, and we treat the composites as descriptive summaries of the music, never as measurements of listeners. Extraction procedure. Each corpus file is parsed with music21 [5] into a flat note stream. Non-note events and files with fewer than eight notes are discarded, the latter because the interval and entropy statistics are unstable on very short fragments. For each retained melody we compute the features of §5 in a single pass, attach the country, region, and repertoire labels, and record a melody title for deduplication. Within each country we then remove duplicate titles, shuffle with a fixed seed, and keep the first 400 survivors, so the per-country sample is a reproducible random subset rather
than the collection’s arbitrary file order. The procedure is deterministic end to end, and re-running it reproduces the melody table byte for byte. Statistics. For each feature we test whether the country distributions differ using the Kruskal-Wallis H test, a non-parametric one-way test appropriate for the skewed, unequal-sized country samples, which does not assume normality or equal variances. We test the seven structural features and the computed minor-mode indicator, and because the effects are extremely strong the conclusion is insensitive to any reasonable multiple-comparison correction, so we report the raw p values and note that all survive a Bonferroni correction across the eight tests by many orders of magnitude. We summarise the between-country structure with a Euclidean distance matrix on the standardised feature means and a principal-component projection of the same matrix. To quantify how much geographic information the features carry we run the supervised benchmark of §8, and to measure the size and stability of the differences we compute Cliff’s delta effect sizes and bootstrap confidence intervals. To test the inferential half of the folk intuition, we correlate the country-level musical-affect measures with the two external indices using the Spearman rank correlation, which is robust at the small country-level sample size. Crucially, the tonal valence cue is correlated with the indices only over the European tonal-idiom countries, because the major-minor construct does not apply to the pentatonic melodies of the Chinese subset, and including them would inject a category error into the valence analysis.
5. Feature Engineering and Mathematical Definitions We define every feature formally so the pipeline is reproducible from the equations alone. Let a melody be a sequence of N notes with MIDI pitch numbers p1 , . . . , pN and note durations d1 , . . . , dN in quarter-length units. The melodic intervals are Ii = pi+1 − pi for i = 1, . . . , N − 1. Register and range. Mean pitch and pitch range are N
p̄ =
1 X pi , N i=1
R = max pi − min pi . i
i
(1)
Interval statistics. The mean absolute interval, the leap ratio (the fraction of intervals wider than a whole tone, i.e. |Ii | > 2 semitones), and the ascending ratio are N −1 1 X |Ii |, N − 1 i=1
(2)
N −1 1 X ⊮[ |Ii | > 2 ] , N − 1 i=1
(3)
I¯ =
L=
N −1 1 X A= ⊮[ Ii > 0 ] , N − 1 i=1
(4)
TABLE 1. P ER - COUNTRY MELODIC PROFILE OVER THE ANALYSED MELODIES . n IS THE DEDUPLICATED , CAPPED SAMPLE SIZE . M INOR IS THE COMPUTED MINOR - MODE FRACTION ( NOISY, AND NOT A VALID VALENCE CUE FOR NON - TONAL IDIOMS SUCH AS THE C HINESE SUBSET ). L EAP IS THE MEAN FRACTION OF INTERVALS LARGER THAN A WHOLE TONE , I NTERVAL IS THE MEAN ABSOLUTE MELODIC INTERVAL IN SEMITONES , AND A ROUS . AND VAL . ARE THE STANDARDISED AROUSAL AND VALENCE COMPOSITES .
Country
Region
Alsace Austria Britain China Czechia Germany Hungary Iran Israel Lorraine Netherlands Russia Switzerland Turkey USA Yugoslav
Western Europe Central Europe Western Europe East Asia Eastern Europe Central Europe Eastern Europe Middle East Middle East Western Europe Western Europe Eastern Europe Central Europe Middle East North America Southeast Europe
n
Minor
Leap
Interval
Arous.
Val.
89 102 400 400 43 400 42 29 94 65 83 37 92 400 6 111
0.07 0.10 0.20 0.38 0.26 0.19 0.29 0.66 0.58 0.41 0.47 0.11 0.10 0.53 0.17 0.21
0.32 0.38 0.36 0.44 0.29 0.30 0.31 0.07 0.29 0.22 0.21 0.29 0.34 0.16 0.38 0.29
2.23 2.60 2.57 2.77 2.20 2.17 2.12 1.10 2.26 2.01 1.93 2.23 2.37 1.69 2.27 2.20
+0.05 +0.56 +1.02 +1.24 -0.32 -0.14 -0.24 -2.20 +1.31 -0.35 -0.66 -0.07 +0.35 +0.52 -0.38 -0.69
+0.18 +0.44 +1.24 +0.27 -0.13 -0.02 -0.23 -2.42 -0.17 -0.30 -0.30 +0.32 +0.76 +0.20 -0.13 +0.30
where ⊮[·] is the indicator function. The leap ratio is the single most discriminative feature in our results, and it is a pure contour statistic that does not depend on any tonal frame. P Note density. With total score duration T = i di , note density is ν = N/T , notes per quarter-length. Pitch-class entropy. Let nc be the count of notes whose pitch class is c ∈ {0, . . . , 11} and qc = nc /N the empirical pitch-class distribution. The Shannon entropy H=−
11 X
qc log2 qc
(5)
c=0
measures how evenly the twelve pitch classes are used. A diatonic melody concentrates mass on seven classes and a pentatonic one on five, so H separates scale systems. The Chinese pentatonic subset has one of the lowest entropies in our data. Mode by key-finding. We estimate mode with the Krumhansl-Schmuckler algorithm [22]. Let x be the duration-weighted pitch-class profile of the melody and kmaj , kmin the empirical major and minor key profiles. For each of the twelve rotations ρ of each profile we compute the Pearson correlation P (xc − x̄)(kρ,c − k̄) qP r(x, kρ ) = pP c , (6) 2 2 (x − x̄) (k − k̄) c c c ρ,c and assign the mode of the profile whose best rotation maximises r. We compute this rather than read the stored label because the corpus documentation warns the stored labels are unreliable, and we treat it as valid only for tonal idioms. Standardisation and composites. For a country-level feature mean m over the set of countries, the standardised value is z(m) = (m−µ)/σ with µ, σ the mean and standard
deviation across countries. The arousal and valence composites average the z-scores of features whose affect direction is established in the music-psychology literature [8], [9], [10]. ¯ + z(R) , Arousal = 31 z(ν) + z(I) (7) 1 Valence = 3 z(1 − minor) + z(p̄) + z(A) . (8) We report the raw features alongside the composites, so no conclusion depends on these particular definitions, and we treat both as descriptive summaries of the music. Effect size. For two groups of leap-ratio values a and b, Cliff’s delta [23] is the ordinal dominance #{(i, j) | ai > bj } − #{(i, j) | ai < bj } δ= ∈ [−1, 1], |a| |b| (9) which requires no distributional assumption and is interpreted by the conventional thresholds (negligible, small, medium, large). Rank correlation. For the index analysis we use Spearman’s ρ, the Pearson correlation of the rank-transformed variables, which is robust to monotone nonlinearity and to the small country-level sample.
6. The Geography of Musical Structure The differences are real. Table 2 reports the KruskalWallis tests over the folk countries. Every one of the 7 structural features, and the computed minor-mode indicator, differs across countries at p < 0.001. The leap-related features are the most sharply separated, with the leap ratio reaching p < 10−90 . These are not marginal effects. The melodic surface of folk song varies strongly and systematically with geography, consistent with the regional-variation half of the cross-cultural literature [6]. Region-specific signatures. Table 1 gives the percountry profile. The clearest folk-song signature is China’s.
standardised composite (z)
Arousal Valence
1 0
−1 −2 l s v y y y a a a a e e d n n A Ira urke Israe Chin rman ustri erlan rland ritai ngar echi ussi osla lsac rrain US R z g T A B u A o e z e C L H h Yu it G Sw Net
Figure 1. Standardised arousal and valence composites by country. Israeli klezmer and Chinese folk sit at the high-arousal end and Iran at the low extreme. The composites summarise the music and are not measurements of listeners.
TABLE 2. K RUSKAL –WALLIS TESTS OF CROSS - COUNTRY DIFFERENCES PER FEATURE . E VERY FEATURE DIFFERS ACROSS COUNTRIES AT p < 0.001 .
Feature Mean pitch Pitch range Mean interval Leap ratio Ascending ratio Note density Pitch-class entropy Minor mode
Kruskal–Wallis H 439.7 638.3 453.8 474.7 253.8 471.5 745.7 175.3
p 1.43e − 86 7.07e − 129 1.51e − 89 5.27e − 94 2.16e − 47 2.55e − 93 7.07e − 152 3.93e − 31
Chinese folk melodies in the corpus use markedly larger intervals (mean absolute interval 2.77 semitones, against Germany’s 2.17) and far more leaps (44 percent of intervals, against Germany’s 30 percent), the widest and most leaping of any tradition in the set, giving a high arousal composite of +1.24 standard deviations. This is the recognisable wideleap, pentatonic contour of the tradition, recovered here from the notes alone. The one tradition with a marginally higher arousal composite is the Israeli klezmer corpus, and that fits its character as fast instrumental dance music, an idiom difference we flag rather than read as geography. The Central European countries, by contrast, cluster around a smoother, more stepwise profile. Figure 1 shows the arousal and valence composites by country. The Iranian classical outlier. The added Iranian corpus sits at the opposite extreme from China, with the smallest mean absolute interval in the set (1.10 semitones, against China’s 2.77 and Germany’s 2.17) and the lowest arousal and valence composites. This matches the stepwise, densely ornamented character of the non-metric radif, but it must be read with the caveats of the previous section firmly in mind. The value is partly an artifact of collapsing a microtonal, non-metric classical repertoire into the same twelve-tone interval and metric features used for the folk melodies, and it
rests on 29 pieces. We show Iran because it is a real Middle Eastern symbolic source and its position is informative, not because a folk-versus-classical, microtonal-versus-diatonic contrast can be read as a like-for-like national comparison. Between-country structure. The distance matrix on standardised feature means, shown in Figure 5, places the Iranian classical corpus as the farthest outlier and groups the Central European countries (Germany, Austria, Switzerland) closely together, a structure that matches geographic and cultural proximity without being told about it. The twodimensional principal-component map in Figure 4 shows the same layout at a glance, and the hierarchical clustering of the same distances in Figure 3 makes the grouping explicit. Iran and Turkey, the classical corpora, join last as the two most distant branches. Among the folk traditions the tree then splits into a large stepwise European cluster (the Central, Western, Eastern, and Southeast European countries together) and a separate wide-leap cluster of China, the Israeli klezmer corpus, and Britain. That wide-leap grouping cuts across geography, since it links a Chinese pentatonic idiom, an Ashkenazi dance idiom, and British dance tunes, and it is a useful caution that the clustering tracks melodic style, which is only partly geographic. The recovery of a plausible cultural geography from melodic statistics alone, by three independent methods, is a check that the features capture real stylistic structure rather than noise.
7. Effect Sizes and Stability A significant test says a difference exists, not that it is large or stable. We report both. Table 3 gives Cliff’s delta [23], a non-parametric effect size, for the sharpest folk-country contrasts on the leap ratio. The China-versusEuropean contrasts are large by the conventional thresholds, reaching a delta of 0.90, and 3 of the four contrasts are large, while a contrast between two structurally similar traditions is negligible. The differences are therefore not only significant but substantial where the traditions genuinely differ
F
G
♯
C
F
G♯
♭
E
B♭
G♯
G
G♯
G
G♯
G♯
G♯
♯
♯
♯
♯
F
F
E
F
G♯
G
G
G
G
Dastgah Shour ♯
G
♯
F
F
E
Persian classical (non-metric, microtonal)
E
D
E
F
E
Makam
Turkish classical (microtonal)
G
C
B
G
D
C
G
D
G
E♭
G
D
G
B
D
G
♭
Klezmer, freygish
Jewish, Israel (Ashkenazi origin)
D
A
C
D
D
A
C
D
G
C
A
G
A
C
D
G
Pentatonic folk Han Chinese
C
F
F
F
G
A
G
F
A
A
A
B♭
♭
C
D
D
C
A
F
F
E
D
C
D
C
B♭
A
F
Diatonic folk
♭
B♭
C
West and Central European
C
A
A
East-Slavic folk
♭
E
♯
C♯
E
A
♯
C♯
E
E
E
A
E
♯
C♯
Russian
E
A
♯
C♯
E
E
Magyar folk Hungarian (modal)
Figure 2. Representative opening phrases rendered as staff notation, one real corpus melody per musical type rather than per country, since the melodic idiom is what differs and several countries share a type. Each panel is labelled by its type with the pitch-class names printed above the staff. The types map to countries as follows. Dastgah Shour is the Iranian classical corpus and Makam the Turkish classical corpus. Klezmer, in the freygish mode, is the Israeli row (Beregovski corpus of Ashkenazi Jewish instrumental folk). Pentatonic folk is the Chinese subset. Diatonic folk is the West and Central European group of Germany, Austria, Switzerland, the Netherlands, Britain, Alsace, Lorraine, Czechia, and the United States. East-Slavic folk is the Russian material and Magyar folk the distinctive Hungarian material, each shown by one exemplar, while the Yugoslav South-Slavic Balkan material is analysed on its own. These seven staves illustrate the main idioms, and the statistics treat every country separately rather than by type. Octave is normalised for display by whole octaves so the contour and interval structure, not the absolute register, are what the reader compares, and the pitch-class names printed above each staff match the notehead positions exactly. The Dastgah Shour and Makam staves are the twelve-tone MIDI approximation of what are really microtonal traditions, so their accidentals stand in for quarter-tones the notation cannot show. The statistical signatures are visible in the notes themselves, with the small stepwise motion of Dastgah Shour, the wide leaps of the pentatonic line, and the smoother stepwise motion of the diatonic folk melodies.
7
TABLE 3. C LIFF ’ S DELTA EFFECT SIZES FOR THE SHARPEST FOLK - COUNTRY CONTRASTS ON THE LEAP RATIO . M AGNITUDE LABELS USE THE CONVENTIONAL THRESHOLDS .
feature-space distance
6 5 4 3 2 1 0
Iran
ds ine ce ey ny nd ria sia lav USA hia ary Turk Aust itzerla therlan Lorra Rus Germa Alsa Hung Czec Yugos Sw Ne
a Chin
el in Isra Brita
Figure 3. Average-linkage hierarchical clustering of the country featurespace distances. The Iran and Turkey classical corpora join last as the most distant branches, and among the folk traditions a stepwise European cluster separates from a wide-leap cluster of China, Israeli klezmer, and Britain.
3
China
USA
2
Yugoslav
1
Austria Switzerland
Russia
Alsace
Hungary
PC2 (30% var)
0
Britain
Germany Netherlands
−1
Iran
Czechia Israel
Lorraine
−2 −3
Contrast
Feature
Cliff’s δ
Magnitude
China vs Germany China vs Netherlands Austria vs Netherlands Germany vs Yugoslav
leap ratio leap ratio leap ratio leap ratio
+0.65 +0.90 +0.60 -0.01
large large large negligible
TABLE 4. 95% BOOTSTRAP CONFIDENCE INTERVALS (2000 RESAMPLES ) FOR EACH FOLK COUNTRY ’ S MEAN LEAP RATIO AND MEAN ABSOLUTE INTERVAL . T HE ESTIMATES ARE STABLE AND THE SHARPLY SEPARATED TRADITIONS DO NOT OVERLAP.
Country
Leap ratio 95% CI
Mean interval 95% CI
Alsace Austria Britain China Czechia Germany Hungary Israel Lorraine Netherlands Russia Switzerland Yugoslav
[0.29, 0.34] [0.34, 0.42] [0.35, 0.37] [0.43, 0.45] [0.26, 0.33] [0.28, 0.31] [0.28, 0.35] [0.28, 0.31] [0.20, 0.24] [0.19, 0.23] [0.24, 0.34] [0.31, 0.38] [0.27, 0.32]
[2.12, 2.34] [2.42, 2.79] [2.52, 2.62] [2.71, 2.82] [2.08, 2.33] [2.13, 2.22] [2.01, 2.24] [2.18, 2.34] [1.92, 2.11] [1.86, 2.01] [2.07, 2.39] [2.21, 2.52] [2.11, 2.29]
Turkey
−4 −5 −6
−4
−2 0 PC1 (58% var)
2
4
Figure 4. Principal-component map of the country feature means. Iran and Turkey lie far from the tight cluster of Central European countries, with China, Israeli klezmer, and Britain spread along the wide-leap direction, recovering a plausible musical geography from the notes alone.
and small where they do not, which is what a meaningful measure should show. Table 4 reports 95 percent bootstrap confidence intervals for each folk country’s mean leap ratio and mean absolute interval. The intervals are narrow and, for the sharply separated traditions, do not overlap, so the between-country gaps are stable properties of the samples rather than artifacts of a few melodies. Together the effect sizes and the bootstrap intervals show that the geography of musical structure is a large and stable effect, which makes the subsequent failure to predict population indices all the more pointed. The musical signal is strong, and it still does not carry to the people.
8. Geographic Classification Benchmark The descriptive tests show the features differ by geography. A stronger, more quantitative question is how much
geographic information the features actually carry. Can a classifier recover a melody’s origin from its features alone? We frame this as a supervised benchmark on the folk melodies, excluding the classical corpora so the classifier cannot win on a repertoire or encoding artifact. Setup. Task A is 13-way folk-country classification and Task B is a coarser 6-way region classification. We evaluate a multinomial logistic regression and a random forest against a majority-class baseline and a stratified-random baseline, under stratified five-fold cross-validation, reporting accuracy and macro-averaged F1 so that small countries are weighted equally with Germany and China. The seven structural features of §5 are the only inputs. Result. Table 5 reports the scores. The random forest reaches 55 percent country accuracy against a 20 percent majority baseline, and 62 percent region accuracy against a 32 percent baseline. The models are well above chance, so the features carry real geographic information, but they are far from perfect, which is the honest and expected outcome. Folk idioms overlap, and no seven-number summary determines a melody’s country. Figure 6 shows the region confusion matrix. East Asia is recovered most cleanly, consistent with the distinctive Chinese profile, while the European regions are most often confused with Central Europe, which matches their musical and geographic proximity. The benchmark quantifies the same structure the descriptive statistics and the distance map show, and it does so with a metric that a reader can compare against future work.
8.4
8.8
5.8
7.5
7.5
5.3
9.2
5.6
5.7
6.2
6.1
6.1
5.4
6.0
0.0
3.1
6.1
4.9
5.4
5.2
4.9
4.4
5.1
5.0
5.2
6.5
4.8
4.3
6.5
3.1
0.0
3.5
3.5
3.2
3.1
4.2
2.9
3.8
3.8
3.6
5.1
3.2
3.6
4.8
6.1
3.5
0.0
3.5
2.3
2.6
4.6
3.2
3.4
3.7
3.2
4.0
3.2
4.2
3.4
4.9
3.5
3.5
0.0
1.8
1.9
1.5
3.8
1.1
0.9
1.0
2.2
0.7
1.1
1.9
5.4
3.2
2.3
1.8
0.0
1.4
3.0
2.8
2.3
2.2
1.9
3.0
1.7
2.6
2.4
5.2
3.1
2.6
1.9
1.4
0.0
2.6
3.0
2.4
2.2
1.5
2.5
1.8
2.5
2.5
4.9
4.2
4.6
1.5
3.0
2.6
0.0
4.5
1.9
1.5
1.6
2.1
1.9
0.9
2.7
4.4
2.9
3.2
3.8
2.8
3.0
4.5
0.0
4.3
4.2
3.9
4.9
3.8
4.1
4.9
5.1
3.8
3.4
1.1
2.3
2.4
1.9
4.3
0.0
0.8
1.2
2.1
0.9
1.4
1.5
5.0
3.8
3.7
0.9
2.2
2.2
1.5
4.2
0.8
0.0
1.0
2.0
0.8
1.1
1.7
5.2
3.6
3.2
1.0
1.9
1.5
1.6
3.9
1.2
1.0
0.0
1.7
0.9
1.4
1.6
6.5
5.1
4.0
2.2
3.0
2.5
2.1
4.9
2.1
2.0
1.7
0.0
2.4
2.5
1.6
4.8
3.2
3.2
0.7
1.7
1.8
1.9
3.8
0.9
0.8
0.9
2.4
0.0
1.4
1.8
4.3
3.6
4.2
1.1
2.6
2.5
0.9
4.1
1.4
1.1
1.4
2.5
1.4
0.0
2.6
6.5
4.8
3.4
1.9
2.4
2.5
2.7
4.9
1.5
1.7
1.6
1.6
1.8
2.6
0.0
8
6
feature-space distance
8.1
4
2
0
Ir a Tur n ke y Isr ae C l Ge hina rm an y Sw Aust itz ria Ne erla the nd rla nd Bri s Hu tain ng Cz ary ec h Ru ia s Yug sia osl a Als v a Lo ce rra ine US A
Iran 0.0 Turkey 8.1 Israel 8.4 China 8.8 Germany 5.8 Austria 7.5 Switzerland 7.5 Netherlands 5.3 Britain 9.2 Hungary 5.6 Czechia 5.7 Russia 6.2 Yugoslav 6.1 Alsace 6.1 Lorraine 5.4 USA 6.0
Figure 5. Between-country Euclidean distance on the standardised structural feature means, with each cell annotated. Iran and Turkey are farthest from every other tradition, and the Central European countries cluster tightly together.
1.0
Central Europe
0.64
0.08
0.01
0.01
0.03
0.23
0.8 0.07
0.84
0.00
0.01
0.01
0.06
Eastern Europe
0.61
0.10
0.03
0.00
0.07
0.20
Middle East
0.27
0.11
0.00
0.40
0.00
0.22
Southeast Europe
0.45
0.07
0.01
0.01
0.37
0.09
Western Europe
0.27
0.06
0.00
0.01
0.01
0.65
Task
Model
country
region
0.6
true
East Asia
TABLE 5. G EOGRAPHIC CLASSIFICATION BENCHMARK UNDER STRATIFIED FIVE - FOLD CROSS - VALIDATION . T HE RANDOM FOREST IS WELL ABOVE BOTH BASELINES ON BOTH TASKS , SHOWING THE FEATURES CARRY REAL BUT IMPERFECT GEOGRAPHIC INFORMATION .
0.4
Accuracy
Macro-F1
Majority baseline Stratified random Logistic regression Random forest
0.204 0.120 0.525 0.546
0.026 0.056 0.244 0.292
Majority baseline Stratified random Logistic regression Random forest
0.325 0.248 0.563 0.621
0.082 0.171 0.446 0.504
0.2
0.0
e ia st pe pe pe op As Ea ro ro ro ur st le Eu Eu Eu a d l tE n n E d a s r r i r a e M te nt st he es Ce Ea ut W So
predicted
Figure 6. Row-normalised confusion matrix of the random-forest region classifier. East Asia is recovered most cleanly, and the European regions are most often confused with Central Europe, matching musical and geographic proximity.
Which features carry the geography. Table 6 reports permutation importance for the region random forest. The strongest single predictor is pc entropy (importance 0.33), which is the feature that separates pentatonic from diatonic scale systems, followed by the three interval and density features. This ranking is musically coherent. The model relies most on the features that encode scale system and melodic contour, exactly the properties that distinguish the traditions in the notation of Figure 2, and least on absolute register. The classifier is therefore not exploiting an incidental cue but the same structural axes the descriptive analysis identifies.
TABLE 6. P ERMUTATION IMPORTANCE OF EACH FEATURE FOR THE REGION RANDOM FOREST ( MEAN ± STANDARD DEVIATION OVER TEN REPEATS ). P ITCH - CLASS ENTROPY, WHICH SEPARATES SCALE SYSTEMS , DOMINATES .
density
Central Europe East Asia
leap
Eastern Europe Middle East Southeast Europe Western Europe
Feature Pitch-class entropy Note density Mean pitch Pitch range Leap ratio Mean interval Ascending ratio
Permutation importance 0.335 ± 0.009 0.176 ± 0.007 0.169 ± 0.007 0.138 ± 0.007 0.125 ± 0.004 0.107 ± 0.006 0.083 ± 0.005
range
interval
9. Regional Affect and Structure Profiles Figure 7 presents the standardised structural profile of each region as a polar diagram, and Figure 8 shows the full per-country distributions of the four sharpest features rather than only their means. The polar view makes the regional signatures legible at a glance. East Asia extends furthest on the interval and leap axes, the Middle East (the Israeli klezmer set) extends furthest on note density and range, and the Central European regions sit near the centre with balanced profiles. The violin plots confirm that the betweencountry differences are differences in the whole distribution, not only in the mean, and that the spreads are comparable across countries, so the mean-based comparisons are not driven by a few outliers. Reading the regions. The polar and violin views are folk-only, excluding the Iran and Turkey classical corpora, so each region is a coherent folk signature. East Asia (the Chinese folk subset) has the widest intervals and the most leaps of any region, with a low pitch-class entropy that reflects its pentatonic basis. This is the sharpest and most separable profile in the data. Middle East here is the Israeli klezmer corpus, and it stands out for the highest note density and a wide range, the fast, busy character of instrumental dance music, which is an idiom difference we flag rather than read as national temperament. Central Europe (Germany, Austria, Switzerland) sits near the centre of every axis with a smooth, stepwise, diatonic character, which is why the classifier uses it as the default guess and why the other European regions collapse toward it. Eastern Europe (Hungary, Czechia, Russia) is close to Central Europe with a slightly narrower range. Western Europe (Netherlands, Britain, Alsace, Lorraine) is pulled wider than the other European groups by the British dance tunes, a reminder that a coarse region can average over real within-region variety. Southeast Europe (the Yugoslav material) has the narrowest range and the lowest note density of all. The Iran and Turkey classical corpora are held apart from these folk regions because their separation is as much a repertoire-andencoding difference as a geographic one, which is exactly why we exclude them from the folk statistics.
entropy
ascending
Figure 7. Polar diagram of the standardised structural features by region. Each axis is a feature z-scored across countries and averaged within region. East Asia extends furthest on the interval, leap, density, and range axes. TABLE 7. S PEARMAN RANK CORRELATIONS OF REGIONAL MUSICAL - AFFECT MEASURES WITH TWO VALIDATED NATIONAL INDICES (W ORLD H APPINESS R EPORT LADDER SCORE , H OFSTEDE INDIVIDUALISM ). N ONE IS SIGNIFICANT AT p < 0.05 .
Music feature
Index
Sample
Arousal (structural) Arousal (structural) Valence (tonal cue) Valence (tonal cue) Minor-mode ratio Minor-mode ratio
WHR IDV WHR IDV WHR IDV
all all tonal only tonal only tonal only tonal only
ρ
n
p
-0.10 -0.46 -0.05 +0.31 -0.11 +0.14
10 9 8 7 8 7
0.776 0.213 0.911 0.504 0.799 0.771
10. Dissociation from National Indices The measurement half of the folk intuition holds. Musical affect varies by region. The inferential half does not. Table 7 reports the Spearman correlations of the regional musical-affect measures with the World Happiness Report ladder score and the Hofstede individualism index. None of the 6 correlations is significant at p < 0.05 (0 of 6). The strongest association, between the arousal composite and individualism, is negative and still not significant at the available sample size, and it is in any case a correlation between a property of eighteenth and nineteenth century folk melodies and a twentieth or twenty-first century survey of living populations, which we would not interpret causally even if it were significant. The plain reading is the honest one. A region whose folk songs sound more aroused, or more minor, or higher, is not thereby a region whose people are measurably happier, sadder, or more individualist. The aggregate-level correlation that the ecological inference would require is not present in the data. Figure 9 shows the arousal composite against the happiness score, a scatter with no trend.
leap ratio
6
0.8 0.6 0.4 0.2 0.0
mean interval
4
Isr ae Ch l Ge ina rm an Au y Sw str it ia Ne zerla the nd rla nd Bri s Hu tain ng a Cz ry ec hi Ru a s Yug sia o sl a Als v ac Lo e rra ine
Isr ae Ch l Ge ina rm an Au y Sw str it ia Ne zerla the nd rla nd Bri s Hu tain ng a Cz ry ec hi Ru a s Yug sia o sl a Als v ac Lo e rra ine
2
note density
PC entropy
3
3
2
2
1
1 Isr ae Ch l Ge ina rm an Au y Sw str it ia Ne zerla the nd rla nd Bri s Hu tain ng a Cz ry ec hi Ru a s Yug sia o sl a Als v ac Lo e rra ine
Isr ae Ch l Ge ina rm an Au y Sw str it ia Ne zerla the nd rla nd Bri s Hu tain ng a Cz ry ec hi Ru a s Yug sia o sl a Als v ac Lo e rra ine
0
Figure 8. Per-country distributions of the four sharpest features across the folk melodies. The differences are distributional, not only in the mean (red bars), and the spreads are comparable across countries.
11. Discussion
World Happiness Report score
7.50
Netherlands
7.25 7.00
Israel Switzerland Austria
Czechia
6.75
Germany
Britain
6.50 6.25 Russia
6.00
China
Hungary
5.75
−1.0
−0.5
0.0 0.5 1.0 regional arousal composite (z)
1.5
Figure 9. The regional arousal composite against the World Happiness Report ladder score, one point per country. There is no trend, illustrating the dissociation between a region’s musical affect and its population’s wellbeing.
What the structural differences mean. The strong, significant between-country differences are differences in musical grammar, the distributions of intervals, contours, and densities that a tradition favours. They are real and worth characterising, and they recover known stylistic signatures such as China’s from the symbolic score. They are statements about music. What the null result means. The absence of any significant correlation with national happiness or individualism is the paper’s most important finding, because it is the one most at odds with the intuition we began with. It says that the affective surface of a region’s traditional music is not a readout of the region’s population psychology. This is what one should expect once the ecological fallacy is taken seriously, and it is worth stating as a measured result rather than only as a caution, because the naive version of this study, the one that reports “country X’s music is sad, therefore country X is sad,” would have produced confident claims from the same data by simply not running the correlation and not questioning the mode labels. The mode-validity problem. The computed minormode ratio is the feature most likely to be over-interpreted, and it is the least valid cross-culturally. The major-minor distinction is a feature of the European tonal system. Applying a key-finding algorithm to a Chinese pentatonic melody returns a major or minor label, but that label is an artifact
of forcing a tonal frame onto music that does not use it, not a valence cue. This is precisely why we restrict the valence correlations to the tonal-idiom countries and why we lean on the mode-independent structural features for the crosscultural comparison. A study that read “China’s melodies are 38 percent minor, so Chinese folk music is often sad” would be committing a category error on top of the ecological one. What the benchmark adds. The classification benchmark turns the qualitative claim that the features carry geographic information into a number, and the number is honest in both directions. It is well above chance, so the information is real, and it is well below perfect, so the features do not determine origin. This is the correct middle position. A benchmark that reported near-perfect accuracy would signal a leak, most likely a repertoire or encoding artifact, which is why we excluded the classical corpora, and a benchmark at chance would contradict the significance tests. The confusion structure is itself a finding. The model’s errors are geographically sensible, confusing neighbours rather than distant traditions, which is evidence that the feature space encodes proximity and not noise. Why the strong signal does not transfer. The most striking pairing in the paper is that the geographic signal is large and stable, with large effect sizes, narrow bootstrap intervals, and above-chance classification, and yet it does not predict the population indices at all. These are not in tension. They say that the music carries a great deal of information about where it comes from and very little about how the people there feel, which are different questions. The strength of the first result is what makes the absence of the second informative rather than merely underpowered. It is not that we failed to find a weak signal, it is that a strong musical signal has no counterpart in the well-being data. Relation to music emotion recognition. Our composites are simpler than the learned models used in music emotion recognition [13], and deliberately so, because our claim is about the dissociation between musical affect and population affect, not about maximising an affect prediction. A more powerful affect model would sharpen the description of each region’s music, but it would not change the central result, since the target the ecological inference needs, a population’s disposition, is not a musical quantity at all.
12. Audio Feature Engineering The symbolic analysis works on notated scores, but the account-level application of §13 must work on audio, where notes are not given and must be summarised by signal features. We describe the Music-Information-Retrieval pipeline formally, both because it is the engineering core of the application and because it is the bridge from a waveform to the same affect proxies used on the scores. Let x[n] be a mono audio signal at sample rate fs . We compute the short-time Fourier transform with frame length M and hop H , X[k, t] =
M −1 X n=0
w[n] x[n + tH] e−ȷ2πkn/M ,
(10)
with w a Hann window, and derive per-frame features that we then average over the track. The root-mean-square energy of frame t is s 2 1 X w[n] x[n + tH] , (11) RMS[t] = M n a proxy for loudness and hence arousal. The spectral centroid, P k fk |X[k, t]| C[t] = P , (12) k |X[k, t]| is the amplitude-weighted mean frequency, a proxy for brightness, and the spectral rolloff is the frequency below which a fixed fraction of the spectral energy lies. The zerocrossing rate counts sign changes of x[n] per frame, rising with noisiness. Tempo is estimated from the autocorrelation of an onset-strength envelope. Pitch content is summarised by the twelve-dimensional chroma vector, X χc [t] = |X[k, t]|, (13) k∈Kc
where Kc is the set of frequency bins mapping to pitch class c, and mode is estimated by correlating the time-averaged chroma against the Krumhansl-Schmuckler profiles as in §5. We implement all of this with librosa [24]. The audio arousal proxy averages the z-scored energy, brightness, and tempo, and the audio valence proxy averages the z-scored brightness and major-mode indicator, mirroring the symbolic composites so that a score and its rendering yield comparable affect summaries. We validated the pipeline end to end on audio rendered from the corpus melodies. The ordering of the rendered per-country audio on the audio arousal proxy matches the ordering of the same melodies on the symbolic composite, which confirms the two representations agree. Popular-song audio by country. The corpus is traditional folk and art music, not what people stream today. To show the same engine on contemporary material, we ran it on a small set of popular-song recordings organised by country, 15 tracks across 6 countries. This is a convenience sample, not a representative one. Some countries contribute a single file, two of the files are multi-song compilations rather than single tracks, and the arousal and valence proxies are z-scored within this small set, so the numbers describe these particular recordings and nothing more. Table 8 reports the per-country summary and Figure 10 plots it. One observation survives the small sample and one cautionary non-result stands out. The recordings are uniformly faster and brighter than the corpus, as expected for produced popular music with a rhythm section. The valence ordering, by contrast, does not carry over from the symbolic corpus to the popular audio for the 6 countries present in both. Iran is low on valence in both domains, but Turkey is the lowest of all in the popular audio while sitting mid-range in the symbolic corpus, the opposite of a match, and the other countries reshuffle between the two. With one to three tracks per country, these audio estimates are far too noisy to read
TABLE 8. P OPULAR - SONG AUDIO SUMMARY BY COUNTRY OVER THE USER - PROVIDED RECORDINGS . n IS THE NUMBER OF ANALYSED FILES , M AJOR RATIO IS THE FRACTION ESTIMATED AS MAJOR MODE FROM CHROMA , AND A ROUS . AND VAL . ARE AUDIO AFFECT PROXIES Z - SCORED WITHIN THIS SET. A SMALL CONVENIENCE SAMPLE THAT DESCRIBES THE RECORDINGS , NOT LISTENERS .
Country
n Tempo Centroid Major Arous. (BPM) (Hz) ratio (z)
China Germany Iran Israel Turkey USA
2 1 4 2 3 3
0.50
2315 2178 2023 2547 1776 1986
0.50 1.00 0.00 0.50 0.00 1.00
Popular-song audio by country arousal valence
0.25 0.00
−0.25 −0.50 −0.75 China Germany Iran Israel Turkey USA (n=2) (n=1) (n=4) (n=2) (n=3) (n=3)
+0.26 +0.28 -0.05 +0.27 -0.31 -0.08
Val. (z)
Country
Track
Peak Quiet Loud pos. third third
+0.37 +0.72 -0.48 +0.63 -0.76 +0.50
China China Germany Iran Iran Iran Iran Israel Israel Turkey Turkey Turkey USA USA USA
Andy Lau BEST OF C POP Apache 207 Wenn das Alireza Talischi Gha Bikalam Turkish Eshg Erfan Tahmasbi Khoda Mohsen Yeganeh Behet Jewish Music 2025 Is Prod By Eli Keshet S Bar Mano Glpembe Mustafa Ceceli Gl Re TARKAN mark Official Morgan Wallen Im A L Ram Jam Black Betty dead prez Hip Hop
0.83 0.45 0.36 0.38 0.60 0.69 0.26 0.70 0.93 0.52 0.94 0.20 0.29 0.45 0.29
Valence, symbolic corpus vs popular audio valence (z, within each domain)
audio affect proxy (z)
0.75
127 123 120 127 142 126
0.5 0.0
−0.5 −1.0 −1.5 −2.0 −2.5
symbolic corpus popular audio China Germany Iran
TABLE 9. W ITHIN - SONG SHAPE SUMMARY FOR EACH POPULAR RECORDING . P EAK POS . IS THE NORMALISED POSITION OF THE ENERGY PEAK (0 AT THE START, 1 AT THE END ), AND Q UIET AND L OUD ARE WHICH THIRD OF THE SONG IS THE SOFTEST AND THE LOUDEST. T HE PEAK POSITION VARIES WIDELY AND IS AS OFTEN AT THE END AS IN THE MIDDLE , SO THE SONGS SHARE NO COMMON EMOTIONAL ARC .
start start start start end end end start start start start end end start end
end end end middle middle middle middle end middle middle middle middle middle middle middle
Israel Turkey USA
Figure 10. Left, arousal and valence audio proxies for the popular-song recordings by country, with the track count n per country. Right, valence for the 6 countries present in both the symbolic corpus and the popular set, computed within each domain. The two orderings do not match. Iran is low in both, but Turkey is lowest in the audio while mid-range in the corpus, so the small audio sample does not reproduce the corpus ordering. Proxies describe the audio, not the listener.
a cross-domain mapping into, and we do not. If anything the mismatch reinforces the paper’s caution. A handful of modern recordings is not a proxy for a tradition, let alone for a population. We present the audio only as an illustration of the engine on real modern material, not as a population measurement. Within-song trajectories carry no shared national arc. A natural next question is whether songs from a country share a temporal emotional shape, for example a slow start, an energetic middle, and a calm close. We tested this directly by computing the full within-song arousal trajectory of each of the 15 popular recordings over twelve-second windows, shown as a small-multiple grid in Figure 11 and summarised in Table 9. The result is a caution rather than a finding. The only feature common to most tracks is a soft intro, with the quietest third at the start in 9 of 15 recordings, and that is a generic production convention rather than a geographic signal. The location of the energy peak is not shared at all. It ranges from 0.20 to 0.94 of the way through the song across the set, falling near the end as often as in the middle, and two of the recordings run the opposite way, beginning bright and ending dark. We report this to guard against a tempting artifact. Summarising a song by three fixed windows forces one of them to look like a peak, so a shared arc can appear where none exists. The defensible cross-domain signal in
this work is the overall affect level of Figure 10, not any within-song shape.
13. Applying the Audio Pipeline to Consented Listening Analytics The same audio pipeline runs on a user’s own listening history, and we release it as a working tool for consented self-analysis. Given a timestamped listening log that a user can export for their own account, and audio the user is entitled to analyse, the engine computes the MIR features of §12 per track and aggregates them into per-account summaries, including mean valence and arousal, a listeningby-hour histogram, and a within-song trajectory. It does not scrape third parties and does not use streaming affect endpoints, because the major services no longer provide them for new applications, so affect is computed only from audio the user owns. A real single-track case study. To show the audio pipeline on genuine material rather than only synthesised demonstration audio, we ran the per-track analysis on a full Persian song, a six-minute recording provided for analysis. The engine reports a global tempo near ninety-six beats per minute, a moderate energy, and a bright spectrum, and it segments the track into fifteen-second windows to produce the within-song affect trajectory of Figure 12. The trajectory recovers real structure. An energetic peak near the three-andthree-quarter minute mark and a sharp calming at the close as the piece resolves. This is a genuine, reproducible read of a real recording, and it is exactly the per-track view the account method aggregates over a listening history. It also carries the paper’s caveat visibly. The computed mode is a Western-tonal reading of a Persian song, an approximation
China: Andy Lau
arousal (z)
2
1
1
0
0
0
−1
−1
−1
−2 0.0
arousal (z)
0.5
1.0
Iran: Alireza Talischi Gha
2
1.0
Iran: Bikalam Turkish Eshg
0.0
1
0
0
0
−1
−1
−1
−2 0.5
1.0
Iran: Mohsen Yeganeh Behet
0.5
1.0
Israel: Prod By Eli Keshet S
0.0
1
1
0
0
0
−1
−1
−1
−2 0.5
1.0
Turkey: Bar Mano Glpembe
2
0.5
1.0
Turkey: Mustafa Ceceli Gl Re
0.0
1
1
0
0
0
−1
−1
−1
−2 0.5
1.0
USA: Morgan Wallen Im A L
2
0.5
1.0
USA: Ram Jam Black Betty
0.0
1
1
0
0
0
−1
−1
−1
−2 0.5
song position
1.0
0.5
1.0
USA: dead prez Hip Hop
2
1
0.0
1.0
−2 0.0
2
−2
0.5
Turkey: TARKAN mark Official
2
1
0.0
1.0
−2 0.0
2
−2
0.5
Israel: Jewish Music 2025 Is
2
1
0.0
1.0
−2 0.0
2
−2
0.5
Iran: Erfan Tahmasbi Khoda
2
1
2
arousal (z)
0.5
1
0.0
Germany: Apache 207 Wenn das
−2 0.0
2
−2
arousal (z)
2
1
−2
arousal (z)
China: BEST OF C POP
2
−2 0.0
0.5
song position
1.0
0.0
0.5
song position
1.0
Figure 11. Within-song arousal trajectories for the 15 popular recordings, one panel per track, over the whole song (horizontal axis is normalised song position, zero to one). The dashed line marks each song’s energy peak. The peak position varies widely across tracks, from 0.20 to 0.94, so there is no shared national arc. The only common feature is a soft intro, a generic production convention. Proxies describe the audio, not the listener.
within-song composite (z)
Arousal Valence
1 0
−1 −2 0
1
2
3 4 time (min)
5
6
Figure 12. Within-song affect trajectory of a real six-minute Persian recording, computed by the audio pipeline over fifteen-second windows. The composites summarise the audio over time, not the listener’s mood. An energetic peak appears near the three-and-three-quarter minute mark and the piece calms sharply at the close.
of the same kind the corpus analysis flags for the Iranian and Turkish material. Two limits define the tool’s scope, and we build both in. The affect proxies describe the audio, not the listener, and the corpus-level null result of this paper is the direct warrant for that restriction. If regional musical affect does not track a population’s well-being, there is no basis for reading an individual’s affect off their playlist. The tool is for consented self-analysis, not for profiling others.
14. Ethics and the Ecological Fallacy This paper deliberately does not, and on its evidence cannot, assign psychological or moral traits to national or ethnic groups. Statements of the form “the people of country X are sad” or “selfish” are stereotypes, and they are not supported by, and do not follow from, any measurement of that country’s music. The corpus measures music, the external indices measure populations, and the central empirical result of the paper is that the two do not correlate. We frame the work this way not as a disclaimer but as its point. The honest handling of a music-and-people question is to measure the music, to test the inference against validated population data, and to report the inference’s failure when it fails.
15. Limitations and Threats to Validity Corpus representativeness. The Essen collection is a curated set of traditional melodies weighted toward Germany and China, not a representative sample of each country’s music, and certainly not of contemporary listening. Our capping and deduplication reduce but do not remove the imbalance, and the smaller countries rest on tens of melodies. Temporal mismatch. The melodies are historical, while the World Happiness Report and Hofstede indices describe
modern populations. Even a significant correlation would therefore be difficult to interpret, which is a further reason we treat the index analysis as a test to reject an inference rather than as a positive estimate of one. Computed mode is noisy. Key-finding on short monophonic folk melodies is imperfect, and the collection warns its stored mode labels are unreliable. We mitigate by computing mode consistently and by not resting the crosscultural comparison on it, but the minor-mode ratios should be read as noisy and, for the non-tonal idioms, as not valid valence cues at all. Small country-level sample. The index correlations have at most eight countries, so their statistical power is low. We therefore do not read the non-significant results as strong evidence of exactly zero association, but as an absence of the clear, strong association the folk inference would require, which is the claim we set out to test.
16. Future Work Several extensions would deepen the study without changing its honest core. On the data side, the geographic coverage is uneven and Europe-and-China heavy, and adding further public symbolic corpora, such as the Meertens Dutch collection, the Session Irish corpus, or Arab-Andalusian material, would broaden the map and let the classification benchmark be evaluated on a more balanced label set. Handling microtonality directly, rather than through the twelvetone MIDI approximation the Iranian and Turkish corpora inherit, would let those traditions enter the quantitative comparison on their own terms rather than only descriptively. On the modelling side, a learned music-emotion model would give a richer affect description than our transparent composites, and a sequence model of melodic expectation would let the entropy and contour features be replaced by a principled predictive measure. None of these changes the central dissociation result, because the population indices the ecological inference needs are not musical quantities, but each would sharpen the description of what the music itself does. On the application side, a prospective study that collected consented listening histories with a validated wellbeing instrument at the individual level would test the same inference where it could in principle hold, at the level of the person rather than the population, which is the only level at which a music-and-mood link is even coherent to ask about.
17. Conclusion The geography of musical affect is real and measurable. Across 2393 melodies from 16 countries, every structural feature we tested differs by country at p < 0.001 among the folk traditions, and regional signatures such as China’s wide-leap, high-activity profile and Iran’s stepwise, lowinterval profile emerge from the symbolic scores alone. The inference from that music to the temperament of the people, however, does not hold. The regional musical-affect measures do not significantly predict national happiness or
individualism, in 0 of 6 tests. Music varies by place, and that variation is worth studying on its own terms, but it is not a measurement of how the people of a place feel, and treating it as one is a fallacy that the data itself declines to support.
[9]
Appendix A. Full Feature Table
[11] J. A. Russell, “A circumplex model of affect,” Journal of Personality and Social Psychology, vol. 39, no. 6, pp. 1161–1178, 1980.
Table 10 gives every structural feature’s mean and standard deviation for every country, so the aggregate profiles of the body can be inspected in full. The classical corpora are shown for completeness with the caveats of §3.
Appendix B. Reproducibility The pipeline is a set of scripts over the public Essen Folksong Collection [2] parsed with music21 [5]. Feature extraction deduplicates by title, caps each country at 400 melodies with a fixed seed, and writes one row per melody. Comparison aggregates by country, runs the Kruskal-Wallis tests, and builds the distance matrix. The index step correlates the country-level measures with the two published indices. Every number in this paper is re-derived from the resulting data files by a checker that also asserts the two central claims as fail-closed invariants, that all cross-country structural differences are significant and that no externalindex correlation is, and the build fails if either ceases to hold. We release the scripts and the derived data tables. We do not redistribute the copyrighted melodies themselves, which are obtained directly from the collection.
References [1]
[2]
W. S. Robinson, “Ecological correlations and the behavior of individuals,” International Journal of Epidemiology, vol. 38, no. 2, pp. 337–341, 2009, reprint of the 1950 American Sociological Review article. H. Schaffrath, “The Essen folksong collection in Kern format,” 1995, d. Huron (ed.). Center for Computer Assisted Research in the Humanities (CCARH), Menlo Park, CA.
[3]
J. F. Helliwell, R. Layard, J. D. Sachs, J.-E. D. Neve, L. B. Aknin, and S. Wang, “World happiness report 2025,” Wellbeing Research Centre, University of Oxford, Tech. Rep., 2025.
[4]
G. Hofstede, Culture’s Consequences: Comparing Values, Behaviors, Institutions and Organizations Across Nations, 2nd ed. Sage Publications, 2001.
[5]
M. S. Cuthbert and C. Ariza, “music21: A toolkit for computer-aided musicology and symbolic music data,” in International Society for Music Information Retrieval Conference (ISMIR), 2010.
[6]
P. E. Savage, S. Brown, E. Sakai, and T. E. Currie, “Statistical universals reveal the structures and functions of human music,” Proceedings of the National Academy of Sciences, vol. 112, no. 29, pp. 8987–8992, 2015.
[7]
S. A. Mehr, M. Singh, D. Knox et al., “Universality and diversity in human song,” Science, vol. 366, no. 6468, p. eaax0868, 2019.
[8]
K. Hevner, “Experimental studies of the elements of expression in music,” American Journal of Psychology, vol. 48, no. 2, pp. 246– 268, 1936.
L. Gagnon and I. Peretz, “Mode and tempo relative contributions to “happy-sad” judgements in equitone melodies,” Cognition and Emotion, vol. 17, no. 1, pp. 25–40, 2003.
[10] P. N. Juslin and P. Laukka, “Communication of emotions in vocal expression and music performance: Different channels, same code?” Psychological Bulletin, vol. 129, no. 5, pp. 770–814, 2003.
[12] T. Eerola and J. K. Vuoskoski, “A comparison of the discrete and dimensional models of emotion in music,” Psychology of Music, vol. 39, no. 1, pp. 18–49, 2011. [13] Y.-H. Yang and H. H. Chen, “Machine recognition of music emotion: A review,” ACM Transactions on Intelligent Systems and Technology, vol. 3, no. 3, pp. 1–30, 2012. [14] R. Panda, R. Malheiro, and R. P. Paiva, “Musical texture and expressivity features for music emotion recognition,” in International Society for Music Information Retrieval Conference (ISMIR), 2018. [15] D. Huron, “Sweet anticipation: Music and the psychology of expectation,” MIT Press, 2006. [16] D. Temperley, “Music and probability,” MIT Press, 2007. [17] E. Foxley, “Nottingham music database,” ABC and MIDI transcriptions of British and Irish folk tunes, 2011, compiled at the University of Nottingham; widely redistributed dataset of traditional melodies. [18] M. Kanani, “ShourCorpus: A symbolic corpus of non-metric iranian classical music (Dastgah Shour),” https://github.com/maziarkanani/ ShourCorpus, 2024, associated corpus paper: Radif Corpus, arXiv:2507.10456. [19] M. K. Karaosmanoğlu, “A Turkish makam music symbolic database for music information retrieval: SymbTr,” in International Society for Music Information Retrieval Conference (ISMIR), 2012. [20] M. Beregovski, “Jewish instrumental folk music,” Klezmer melodies transcribed to Humdrum kern by the Mode-in-Klezmer project, 2024, corpus of Ashkenazi Jewish instrumental folk music; collected by Beregovski (1892–1961), digital edition at shanahdt.github.io/mode in klezmer. [21] J. Malin and D. Shanahan, “Modes in klezmer music: A corpus study based on beregovski’s jewish instrumental folk music,” Music Theory Online, vol. 31, no. 3, 2025. [22] C. L. Krumhansl, Cognitive Foundations of Musical Pitch. University Press, 1990.
Oxford
[23] N. Cliff, “Dominance statistics: Ordinal analyses to answer ordinal questions,” Psychological Bulletin, vol. 114, no. 3, pp. 494–509, 1993. [24] B. McFee, C. Raffel, D. Liang, D. P. W. Ellis, M. McVicar, E. Battenberg, and O. Nieto, “librosa: Audio and music signal analysis in Python,” in Proceedings of the 14th Python in Science Conference (SciPy), 2015.
TABLE 10. F ULL PER - COUNTRY STRUCTURAL FEATURE MEANS WITH STANDARD DEVIATIONS IN PARENTHESES . R EGISTER AND INTERVALS IN SEMITONES , DENSITY IN NOTES PER QUARTER - LENGTH , ENTROPY IN BITS . Country Alsace Austria Britain China Czechia Germany Hungary Iran Israel Lorraine Netherlands Russia Switzerland Turkey USA Yugoslav
Pitch
Range
Interval
Leap
Asc
Density
Entropy
68.59 (3.62) 68.58 (3.68) 72.13 (2.33) 70.79 (4.26) 68.95 (3.22) 68.09 (3.86) 69.22 (2.49) 64.39 (5.58) 71.20 (1.41) 69.13 (1.68) 68.65 (3.09) 68.62 (2.78) 69.14 (3.60) 74.13 (1.71) 67.67 (3.35) 68.61 (2.53)
12.60 (3.00) 13.87 (4.12) 16.89 (3.25) 15.38 (3.32) 11.35 (2.58) 12.69 (2.78) 12.29 (3.18) 10.65 (3.38) 16.25 (4.08) 12.41 (2.13) 11.29 (2.60) 11.78 (2.92) 12.56 (3.01) 16.55 (3.62) 10.83 (2.03) 9.73 (3.17)
2.23 (0.54) 2.60 (0.94) 2.57 (0.54) 2.77 (0.51) 2.20 (0.44) 2.17 (0.48) 2.12 (0.37) 1.10 (0.32) 2.26 (0.40) 2.01 (0.40) 1.93 (0.36) 2.23 (0.53) 2.37 (0.75) 1.69 (0.26) 2.27 (0.64) 2.20 (0.47)
0.32 (0.12) 0.38 (0.18) 0.36 (0.12) 0.44 (0.11) 0.29 (0.12) 0.30 (0.12) 0.31 (0.11) 0.07 (0.05) 0.29 (0.08) 0.22 (0.08) 0.21 (0.10) 0.29 (0.14) 0.34 (0.15) 0.16 (0.07) 0.38 (0.15) 0.29 (0.15)
0.35 (0.08) 0.39 (0.09) 0.43 (0.06) 0.38 (0.07) 0.35 (0.08) 0.36 (0.08) 0.34 (0.07) 0.26 (0.05) 0.37 (0.06) 0.36 (0.08) 0.38 (0.08) 0.37 (0.09) 0.41 (0.10) 0.35 (0.05) 0.35 (0.05) 0.39 (0.10)
1.44 (0.38) 1.42 (0.40) 1.44 (0.39) 1.76 (0.42) 1.27 (0.39) 1.25 (0.47) 1.26 (0.35) 0.36 (0.04) 2.21 (0.60) 1.23 (0.34) 1.16 (0.40) 1.45 (0.37) 1.65 (0.45) 1.83 (0.50) 1.22 (0.35) 1.13 (0.34)
2.47 (0.25) 2.50 (0.26) 2.67 (0.18) 2.27 (0.18) 2.46 (0.20) 2.48 (0.25) 2.38 (0.29) 2.31 (0.31) 2.66 (0.17) 2.51 (0.19) 2.48 (0.18) 2.39 (0.26) 2.46 (0.34) 2.88 (0.20) 2.28 (0.34) 2.24 (0.25)