- The paper presents the MPB Corpus, a symbolic dataset The dataset covers 500 Brazilian Popular Music compositions with melodic results, harmonic relations, and metric analyses characterized the genre.
- Melodic rhythm analysis shows high baseline syncopation on reserve cadences and determined CMI values varying by composer styled cadences.
- MPB Corpus enables statistical discrimination between composers based on rhythm with evidence that stylistic information is carried across melodic and harmonic dimensions of songs.
The MPB Corpus, introduced by Almada, Carvalho, and Martins (2608.13842), is a symbolic dataset of 500 Brazilian Popular Music (MPB) compositions encoded across four musical parameters: melodic contour, melodic rhythm, harmony, and melody–harmony relationships. The corpus covers 50 songs each by ten composers—Tom Jobim, Ivan Lins, Chico Buarque, Edu Lobo, Caetano Veloso, Djavan, João Bosco, Milton Nascimento, Gilberto Gil, and Rita Lee—and contains 8,426 c- and r-words, functional information for 17,053 chords, and melody–harmony relations for 23,447 notes. The authors position it as the most comprehensive and detailed dataset to date for computational musicology of Brazilian music, a claim that is plausible given that prior work on Brazilian repertoires has addressed only single dimensions (e.g., harmony in choro, rhythm in samba).
Motivation and research question
The paper frames the dataset around the notion of an "MPB Common Practice," proposed in analogy with Tymoczko's extended common practice. The motivating question is which stylistic markers—in rhythmic and melodic contour, harmony, and melody–harmony interaction—function as shared norms across MPB composers, and which diverge to characterize individual styles. The ten composers were selected on the basis of critical consensus, commercial success, and, crucially, the availability of high-quality published songbooks, since the analytical pipeline depends on reliable scores.
A notable methodological stance is the restriction to the "nominal form" of compositions: only main melodic sections are analyzed, excluding introductions, interludes, codas, and literal repetitions. This prioritizes essential structural material over realized arrangements, but it also means the corpus does not capture performance practice or variation across recordings.
Analytical models
Genera of Chord Types (GCT). The harmonic model operates at a "semantic" level, defining 10 protochords (basic triads and seventh chords) as genera, from which 161 chord types are derived via recursive ADD, SUB, and ALT transformations under seven constraining rules. Chord types receive both alphanumeric labels (roughly conventional) and genealogical notation encoding derivative lineage within a genus. The distribution is highly uneven: the dominant genus Y contains 59 members, reflecting its combinatorial richness in tensions and alterations, while genus w contains only four. A "syntactic" level modeling relationships between chords is explicitly deferred to future work, so the current release encodes chords without their sequential context—a significant limitation for studying harmonic progressions directly.
Melodic Filtering Model (MFM). Melodies are segmented into "words" (analogous to Lerdahl's groups), then pitch content is encoded as sequences of c-letters spanning seven contour gestures (repetition, stepwise motion, arpeggio, leap, ascending/descending), a refinement of Parsons code. Rhythm is encoded as r-letters: the beat is subdivided into 12 units, and inter-onset interval configurations within this window map onto an alphabet of 26 elements, of which letters "a" through "v" suffice for nearly the entire repertoire, with four wildcard letters reserved for exceptional cases.
Two metrics accompany these encodings. The compensated intervallic economy index (CIEI), normalized to [0,1], captures the balance of upward and downward motion within a c-word; values near one indicate compensatory rectification of melodic direction. The countermetricity index (CMI) quantifies syncopation against the metrical grid, a parameter the authors argue is constitutive rather than exceptional in Afro-Brazilian genres.
Melody–harmony relations. Building on chord/scale theory, notes are classified as structural or inflections relative to local (chordal) and global (key) contexts, using the "two semitones" rule for tension eligibility. The NF web visualization arranges these classifications on five concentric orbits (inflections through altered tensions) crossed by twelve chromatic rays, yielding 60 positions whose occupancy profiles characterize a piece's melody–harmony complexity. The melodic anchoring index (MAI) summarizes this as a weighted average of note functions, normalized to [0,1].
The encoding workflow is largely manual: contour extraction is automated from MIDI, but rhythmic encoding, harmonic analysis, and note-function annotation are performed by hand. Harmonic analysis in particular is identified as not reliably automatable with current techniques.
Exploratory findings
The exploratory analysis yields several concrete observations. In melodic rhythm, the r-letter "b" (a single onset at beat start) dominates most corpora, but Rita Lee's metric profile is anomalous: position 7 (the beat midpoint) reaches 48.2%, exceeding even position 1. Positions dividing the beat into thirds (5 and 9) are markedly more frequent in Jobim (7.4% and 7.5%) and Ivan Lins (10.7% and 10.9%), consistent with bossa nova's ternary subdivisions. Countermetric r-letters account for roughly one-third of all occurrences corpus-wide, confirming high baseline syncopation; samba-oriented repertoires (Jobim, Buarque, Bosco, Djavan, Gil) show CMI above 0.53, while jazz- or rock-influenced repertoires (Lins, Lobo, Lee) show different countermetric signatures dominated by letter "e."
In melodic contour, descending stepwise motion ("p") is the most frequent c-letter in eight of ten corpora, prompting the hypothesis that this asymmetry may be characteristic of Brazilian popular melody—an open question the paper flags for further study. Milton Nascimento stands apart, with note repetition dominant by a wide margin, reflecting his characteristically static melodic construction. CIEI values are uniformly high (0.856–0.898) across corpora, supporting a compensatory law in melodic lines, and modal word lengths converge on six to seven attack points, consistent with prior segmentation studies.
Harmonically, genus Y dominates nearly all corpora, reflecting shared tonal syntax and dominant preparation; genus V is disproportionately present in Veloso, Gil, and Lee, consistent with pop-rock adherence, while Nascimento's triad-heavy profile reflects modal syntax. Functional analysis shows I and V dominating everywhere, with chromatic categories distributed idiosyncratically—the authors interpret this as shared diatonic foundations with individualized "spice." For melody–harmony relations, Djavan, Edu Lobo, Ivan Lins, and especially Jobim show elevated tension usage, suggesting a subgroup with more complex melody–harmony interaction; the contrasting NF webs of Rita Lee's "Saúde" (MAI = 0.817) and Jobim's "Angela" (MAI = 0.416) illustrate the metric's discriminative range.
Statistical validation
To demonstrate quantitative use, the authors test whether r-letter distributions discriminate between composers. They define a ratio statistic R comparing mean Jensen-Shannon distance between pieces by different composers to that within composers, and assess significance via permutation testing with 10,000 label permutations. The observed value was Robs​=1.072, exceeded by none of the permuted statistics, yielding p<10−4. This provides strong evidence that rhythmic distributions carry composer-specific stylistic information. The effect size is modest in absolute terms (R close to 1), indicating substantial overlap between composers' rhythmic profiles—which is itself consistent with the paper's thesis of a shared common practice overlaid with individual differentiation—but the statistical signal is unambiguous.
Limitations and open questions
Several limitations are acknowledged explicitly. The annotator pool consists of a single expert (the first author) for the entire corpus, ensuring consistency but precluding inter-annotator reliability assessment. The composer selection is predominantly male, partly because influential female performers (Gal Costa, Elis Regina) are documented primarily through performance, and female composers (Dona Ivone Lara, Maysa, Joyce) lack readily available published scores; the authors state that addressing this requires audio transcription outside the current scope. The melody–harmony layer analyzes only a contiguous excerpt per piece rather than complete works, justified empirically rather than systematically. Melodic segmentation remains the least formalized stage of the pipeline. The syntactic level of harmony—chord-to-chord transitions—is absent from the current encoding, limiting direct study of harmonic progressions. Finally, the exploratory analyses are presented as illustrative rather than conclusive; hypotheses such as the predominance of descending steps in Brazilian melody and the existence of a tension-rich composer subgroup await rigorous testing, including clustering analyses.
Conclusion
The MPB Corpus consolidates roughly five years of methodological development into a publicly available, multi-layer symbolic resource for Brazilian popular music, accompanied by novel encoding schemes (GCT, MFM), purpose-built metrics (CIEI, CMI, MAI), and reproducible Python tooling hosted on Zenodo and GitHub. Its demonstrated capacity to statistically separate composers on rhythmic grounds, combined with planned expansions—to approximately fifty composers, a 150-song control corpus of choro, samba, and jazz, and sequential/syntactic extensions—positions it as a foundation for quantitative studies of style, influence, and historical change in a repertoire that has remained largely outside digital musicology.