Source-linked AI summary
Automatic Analysis and Influence of Hierarchical Structure on Melody, Rhythm and Harmony in Popular Music
Shuqi Dai, Huan Zhang, Roger B. Dannenberg
TL;DR
Music-structure analysis must explain how hierarchical levels shape musical organization beyond repetition alone. This paper extracts repeated phrases and infers sections from MIDI, then tests their relationships with harmony, melody, and rhythm. The resulting structures achieve 93% accuracy against human labels and reveal level-specific musical interactions, while section detection and style coverage remain limited.
Problem
Existing music-structure analysis often ignores hierarchy, leaving unclear whether phrases and sections play different roles in organizing harmony, melody, and rhythm.
Method
The paper extracts repeated phrase structures from MIDI, infers higher-level sections, and analyzes musical elements at both hierarchical levels.
Results
93% accuracy against human labeling accompanies evidence that harmony, melody, and rhythm behave differently at positions within hierarchical structure.
Takeaways & Limitations
Hierarchical structure can inform music analysis, theory evaluation, music generation, imitation, and assessment of structural conventions.
Takeaways & Limitations
Section detection relies on non-melodic phrases that are not present in all styles, and the study calls for comparisons with Western pop and other datasets.
Abstract
from arXiv · showhide
Repetition is a basic indicator of musical structure. This study introduces new algorithms for identifying musical phrases based on repetition. Phrases combine to form sections yielding a two-level hierarchical structure. Automatically detected hierarchical repetition structures reveal significant interactions between structure and chord progressions, melody and rhythm. Different levels of hierarchy interact differently, providing evidence that structural hierarchy plays an important role in music beyond simple notions of repetition or similarity. Our work suggests new applications for music generation and music evaluation.
1 Introduction
The paper develops hierarchical repetition analysis because music structure spans phrases and sections, yet the distinct roles of these levels are not well formalized. It extracts phrase and section structures to study how hierarchy relates to harmony, melody, and rhythm.
- Music structure ranges from low-level motives to higher-level phrases and sections, with these levels influencing harmony, melody, and rhythm.
- The study identifies phrases through repeated harmony, melody, and rhythm, then infers higher-level sections from phrase-level structure.
- Its main contributions are a repetition-structure algorithm for phrase and section levels, evidence that musical elements reflect hierarchy, and data-driven features for music theory and MIR.
2 Related Work
Related MIR work addresses musical form, hierarchical chord structure, segmentation, and melody or chord analysis, but these efforts use varied representations and tasks. MIDI offers cleaner rhythmic information and melody separation than audio for this study’s analysis.
- Prior work covers computational form analysis, hierarchical chord sequences, tree structures, and segmentation of classical and popular music.
- Audio segmentation methods use repetition, textural similarity, contrast, timbre, and rhythmical properties.
- MIDI provides more reliable rhythmic information and cleaner melody separation than audio for structure analysis.
- Related research also includes chord recognition, MIDI melody extraction, matching melodic phrases, segmentation evaluation, and music-structure analysis.
3 Phrase-level Structure Analysis
The phrase-level method represents repeated musical segments compactly and selects structures by minimizing Structure Description Length. It searches matched phrase sets with graph-based optimization and achieves high agreement with human-labeled structures, though runtimes vary substantially.
- 3. Phrase-level Structure Analysis: Structure labels alternate phrase letters with measure lengths, distinguishing repeated melodic phrases, non-melodic phrases, near-repetitions, and unmatched phrases.Capital and lowercase labels distinguish melodic and non-melodic phrases, while X and x mark non-repeating phrases.
- 3. Phrase-level Structure Analysis: Matching letters represent repeated measures based on similarity of chords, rhythm onset times, and melodic distance, although songs can have multiple valid structures.
- 3. Phrase-level Structure Analysis: SDL favors compact repeated descriptions, using h · 3 + g · 8 for ABA versus h · 1 + g · 12 for a single longer phrase representation.The manually tuned constants are h = 1.0 and g = 1.3 on training data.
- 3. Phrase-level Structure Analysis: Given melody, chord analysis, and time signature, the algorithm finds a phrase structure with the shortest Structure Description Length.It matches non-overlapping segments of 4–20 measures, merges them into phrase sets, and optimizes the resulting structure.
- 3. Phrase-level Structure Analysis: The optimization finds maximal cliques of matched phrases and then solves an NP-complete maximum weighted clique problem using dynamic programming, A* search, and heuristics.
- 3. Phrase-level Structure Analysis: The full algorithm correctly produced 92% of human-labeled structures, with a 345-second average runtime and 21 seconds for 80% of songs.
4 Hierarchical Structure Exploration
The analysis identifies recurring phrase-level patterns and uses them to characterize a two-level structure of phrases grouped into higher-level sections. Most songs show compact sectional organization, with substantial repetition within and across sections.
- Repeated melodic phrases cover 50% to 90% of the song in most songs.
- Most melodic phrases are four or eight measures long.
- The method defines sections by using sequences of non-melodic and non-matching phrases as boundaries around lower-level repetition segments.
- Most songs contain two or three sections, with each section typically comprising one to six phrases.
- Over 90% of songs have two or three distinct melodic phrases, and each section typically contains one to three distinct melodic phrases.
- 20% of sections exactly repeat the previous section, while 29% repeat its suffix and 18% repeat its prefix.
5 Interactions with Segment Structure
Hierarchical phrase and section positions systematically shape harmony, melody, and rhythm, with distinct interactions at the two structural levels.
- I chords are more likely at phrase beginnings and section endings, whereas V is more common at phrase endings within sections.All differences from the equal-probability background are significant at P < .0001.
- 58% of major-mode section endings are V–I, with V–I transition probabilities of 0.89, 0.84, and 0.94 at phrase, within-section phrase, and section endings versus 0.47 elsewhere.All reported transition differences are significant, and section endings favor I more than phrase endings.
- Scale degree 1 tends to occur at section endings, while scale degree 3 is common at phrase beginnings but not section endings under I chords.
- Whole-or-longer notes comprise 6.4% of phrase endings within sections but 72% of phrase endings at section endings.Phrase beginnings and middles mostly contain short notes, whereas phrase endings favor longer notes.
- Cross-phrase similarity decreases and phrase complexity increases with composition date across seven decades of music.The reported associations are significant at P < .01 and P < .002, respectively.
6 Discussion and Future Direction
The discussion interprets the detected hierarchy as musically meaningful because harmony, melody, and rhythm vary by structural position and align with music-theoretic patterns. The algorithms achieve 93% agreement with human labeling, while manual tuning, limited style coverage, and incomplete interaction analysis remain boundaries.
- Harmony, melody, and rhythm behave differently across hierarchical positions, including half cadences at phrase endings in the middle of sections.The phrase extractor uses repetition and similarity without music-theoretic knowledge, allowing the results to test existing theory.
- 93% accuracy against human labeling supports using the hierarchical repetition algorithms to analyze other MIDI datasets.
- The findings could guide music imitation or generation and evaluate whether songs follow structural conventions.
- Phrase-level parameters are manually tuned, and non-melodic phrases used for section detection are absent in some styles.The authors also call for comparisons with non-pop and Western pop datasets and more robust section indicators.
- The study has only begun characterizing interactions among structure, melody, harmony, and rhythm, leaving variations and contrasting-phrase construction for future work.
7 Conclusion
The study introduces a data-driven hierarchical structure analysis algorithm and applies it to harmony, melody, and rhythm. The released dataset, annotations, and experimental results support future work in music style, analysis, evaluation, and generation.
- The study introduces a hierarchical structure analysis algorithm for examining harmony, melody, and rhythm across multiple structural levels.
- The work provides evidence that hierarchical structure can inform future music style, analysis, evaluation, and generation research.
- The dataset, annotations, and experimental results are released publicly for further use.