Source-linked AI summary
POP909: A Pop-song Dataset for Music Arrangement Generation
Ziyu Wang, Ke Chen, Junyan Jiang, Yiyi Zhang, Maoran Xu, Shuqi Dai, Xianbin Gu, Gus Xia
TL;DR
Arrangement models lack refined datasets for evaluation and practical results. POP909 addresses this gap with 909 professionally created, multi-version piano arrangements aligned to melodies and audio and enriched with annotations. Baseline experiments show the dataset supports generation that captures basic harmonic and rhythmic structure, while chord-label agreement varies across songs.
Problem
Arrangement research lacks suitable datasets combining style-consistent arrangements, precise source alignment, and external labels for controllable generation.
Method
The paper constructs POP909 from professionally arranged and reviewed popular songs, providing aligned MIDI arrangements, audio, melodies, and tempo, beat, key, and chord annotations.
Results
Baseline Transformer generation captures basic harmonic relationships between melody and accompaniment and consistent rhythmic patterns.
Takeaways & Limitations
POP909 supports piano accompaniment generation, audio-based re-orchestration, structural music generation, and cross-modal music generation.
Takeaways & Limitations
Chord annotations are not necessarily consistent between arrangements and original audio; more than 800 songs exceed 75% root-note matching, but some fall below 40%.
Abstract
from arXiv · showhide
Music arrangement generation is a subtask of automatic music generation, which involves reconstructing and re-conceptualizing a piece with new compositional techniques. Such a generation process inevitably requires reference from the original melody, chord progression, or other structural information. Despite some promising models for arrangement, they lack more refined data to achieve better evaluations and more practical results. In this paper, we propose POP909, a dataset which contains multiple versions of the piano arrangements of 909 popular songs created by professional musicians. The main body of the dataset contains the vocal melody, the lead instrument melody, and the piano accompaniment for each song in MIDI format, which are aligned to the original audio files. Furthermore, we provide the annotations of tempo, beat, key, and chords, where the tempo curves are hand-labeled and others are done by MIR algorithms. Finally, we conduct several baseline experiments with this dataset using standard deep music generation algorithms.
1 Introduction
Music arrangement reconstructs and reconceptualizes pieces across conditional generation tasks, but progress is constrained by inadequate datasets. POP909 addresses this gap with professionally created, aligned piano arrangements and annotations for accompaniment and audio-conditioned re-orchestration.
- Music arrangement includes accompaniment generation from lead sheets, audio-conditioned transcription and re-orchestration, and reduction of full scores for fewer instruments.
- Piano arrangement connects lead sheets, audio, and full scores while offering rich musical expression for future automated composition and performance.
- Suitable arrangement datasets require style-consistent re-orchestration, precise alignment with source music, and external labels such as chords and downbeats.
- POP909 contains 909 popular songs with multiple professional piano-arrangement versions aligned to lead melody and original audio, plus tempo, beat, key, and chord annotations.
- The dataset supports piano accompaniment generation conditioned on melody and auxiliary annotations, audio-based re-orchestration, and additional symbolic or expressive-generation tasks.
2 Related Work
Existing music datasets provide varied modalities and annotations, but often lack arrangement-specific structure, alignment, or sufficient scale. POP909 is designed around these requirements for piano arrangement generation.
- 2.1 Modalities of Music Generation: Music-generation datasets span symbolic score generation, performance rendering, and audio synthesis, while POP909 targets arrangement in score and performance modalities.
- 2.2 Existing Datasets: Existing datasets are evaluated by properties including modality, arrangement relevance, alignment, annotations, and dataset size.
- 2.2 Existing Datasets: Lakh MIDI contains 176,581 songs but does not identify lead melody or piano accompaniment tracks, preventing direct use for piano arrangement.
- 2.2 Existing Datasets: Maestro and E-piano provide time-aligned classical piano MIDI and audio, but ambiguous melody boundaries and transcription-based MIDI limit their suitability for arrangement tasks.
- 2.3 Requirements of Datasets for Piano Arrangement: POP909 is designed to provide style-consistent piano tracks, lead melody or audio, rich annotations, time alignment, and considerable size for deep learning.
3 Dataset Description
POP909 combines professionally reviewed piano arrangements of 909 popular songs with multiple versions, expressive three-track MIDI, aligned audio, and beat, chord, and key annotations.
- Dataset Overview: POP909 contains time-aligned piano arrangements of 909 popular songs that maintain the original style and texture, with beat, chord, and key information.
- 3.1 Data Collection Process: Professional arrangers and reviewers iteratively create, assess, discuss, and refine each arrangement to maintain quality.
- 3.1 Data Collection Process: The dataset selects 909 high-quality songs from an initial list of 1000 and retains both qualified final versions and unqualified intermediate versions.
- Dataset Scale: The 909 arrangements total about 60 hours, span 462 artists, and cover releases from the 1950s to around 2010.
- 3.2 Data Content and Format: Each arrangement uses MELODY, BRIDGE, and PIANO MIDI tracks; BRIDGE plus PIANO form the accompaniment arrangement.
- 3.2 Data Content and Format: MIDI arrangements align with original audio through manually labeled tempo curves, and note events include expressive dynamics derived from the audio.
- 3.3 Data Folder Structure: Each song folder contains arrangements, plain-text annotations, and all iterative versions, while a root index supplies song, artist, revision, and related metadata.
4 Annotation Methods
POP909 derives beat, downbeat, chord, and key annotations from MIDI and audio using complementary algorithms and alignment procedures. Beat annotations agree by more than 90% within 100 ms, while chord-root agreement exceeds 75% for more than 800 songs but remains low for some recordings.
- Beat & Downbeat Estimation: Beat positions are estimated from human-labeled tempo curves and accompaniment onset features that indicate beats and downbeats.The MIDI-based procedure uses MIDI onset and velocity features together with tempo information.
- Beat & Downbeat Estimation: The MIDI beat tracker estimates an initial beat, deduces later positions from tempo, infers meter through autocorrelation, and selects the best phase shift.
- Beat & Downbeat Estimation: More than 90% beat-position consistency was obtained within 100 ms, while downbeat annotations achieved 80% agreement between MIDI- and audio-based algorithms.Both extraction results are included in the annotation files.
- Chord Label Extraction: Chord labels are extracted from MIDI and audio, with chord boundaries aligned to beats because popular-music chord changes commonly occur there.MIDI chord templates include pop-oriented triads, sevenths, suspended chords, and sixth chords with inversions where applicable.
- Annotation Files: The annotation files record beat, chord, and key labels, including separate audio- and MIDI-derived beat annotations and two MIDI downbeat extractors.The downbeat extractors assume either simple or compound meter.
- Chord Label Extraction: More than 800 songs have chord-root matching above 75%, but a few fall below 40% because of tuning issues, sound effects, or deliberate chord re-arrangement.Arrangement and original-audio chord progressions are therefore not necessarily consistent.
5 Experiments
The experiments evaluate Transformer-based score generation with POP909 for unconditioned polyphonic music and melody-conditioned piano arrangement. Generated examples capture basic harmonic and rhythmic structure, but remain below state-of-the-art quality and serve primarily as dataset baselines.
- Experimental Setup: Two baseline tasks use Transformers: unconditioned polyphonic music generation and piano accompaniment generation conditioned on melody.The architecture is selected for modeling long-term dependencies in time-series data.
- Polyphonic Music Generation: Polyphonic pieces are encoded as MIDI-like event sequences containing note onsets, offsets, velocity changes, and time shifts.Time-shift tokens are quantized from 1/4 beat to 4 beats, with longer rests or notes represented by multiple tokens.
- Polyphonic Music Generation: The dataset is split into 90% training, 5% validation, and 5% testing subsets.
- Polyphonic Music Generation: The training configuration uses a Transformer with L = 2048, H = 512, h = 6, and N = 6, trained with cross-entropy loss and early stopping.Adam optimization and a warm-up learning-rate schedule are also used.
- Polyphonic Music Generation: The warm-up schedule uses S = 4000, and training results are reported through MIDI-event prediction loss and accuracy.
- Piano Arrangement Generation: Melody-conditioned arrangement generation forces the generated melody to match the input while producing the most likely piano accompaniment.The main melody merges the MELODY and BRIDGE tracks, while the PIANO track supplies accompaniment data.
- Experiment Results: Generated examples show unconditioned polyphonic output above melody-conditioned piano arrangements, with melody in blue and accompaniment in orange.The first 500 MIDI-event tokens provide context before generated output begins.
- Experiment Results: The generated pieces capture basic melody–accompaniment harmonic relationships and consistent rhythms, but their quality remains far below state-of-the-art algorithms.The authors present them as baseline demonstrations of dataset usage.
6 Conclusion
POP909 is a tailored dataset of professional piano arrangements for 909 popular songs, combining aligned MIDI and audio resources with structural annotations. Its collection process emphasizes quality, and the dataset also supports structural and cross-modal music-generation research.
- Conclusion: POP909 contains multiple MIDI versions of professional piano arrangements for 909 popular songs, with tempo curves aligned to original audio recordings.
- Conclusion: The dataset provides tempo, beat, downbeat, key, and chord annotations alongside the arrangements.
- Conclusion: Professional arrangers and reviewers collected the dataset through an interactive process designed to guarantee high data quality.
- Conclusion: Beyond arrangement, POP909 serves as a resource for structural music generation and cross-modal music generation.