[Paper Review] POP909: A Pop-song Dataset for Music Arrangement Generation

Ziyu Wang, Ke Chen+6 more|arXiv (Cornell University)|Aug 17, 2020

Music and Audio Processing33 references51 citations

TL;DR

POP909 provides 909 professional piano arrangements (MIDI) aligned to lead melodies and original audio, with tempo, beat, key, and chord annotations, enabling piano-accompaniment and cross-modal arrangement research using baseline Transformer models.

ABSTRACT

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.

Motivation & Objective

• Provide a high-quality, time-aligned dataset of piano arrangements for pop songs to support arrangement generation tasks.
• Enable piano accompaniment generation conditioned on lead melody and conditioned on audio (re-orchestration) with precise supervisory annotations.
• Facilitate evaluation and development of deep learning models for symbolic music arrangement and cross-modal generation.

Proposed method

• Assemble 909 professional piano arrangements in MIDI aligned to original audio and the vocal lead melody.
• Annotate tempo curves manually and extract beat, key, and chord labels using MIR algorithms.
• Use a two-team arrangement-and-review process to ensure high-quality, style-consistent piano arrangements.
• Represent music as a MIDI-like event sequence with a Transformer model to learn joint distributions of melody and accompaniment.
• Train a GPT-2–style Transformer with relative positional encoding on a MIDI-like event vocabulary to model polyphonic generation.
• Condition piano arrangement generation on the melody by merging MELODY and BRIDGE tracks during inference.

Experimental results

Research questions

• RQ1 How can POP909 be used to learn piano accompaniment generation conditioned on a lead melody or on audio-derived arrangements?
• RQ2 What baseline performance do standard deep generative models (e.g., Transformer) achieve on symbolic arrangement tasks with POP909?
• RQ3 Can the dataset support unconditional symbolic generation, expressive performance rendering, or cross-modal arrangement tasks?

Key findings

• The dataset contains about 60 hours of arrangements across 909 songs spanning from the 1950s to ~2010, with three MIDI tracks per file (MELODY, BRIDGE, PIANO) aligned to audio.
• Tempo curves are manually labeled; beat, key, and chord annotations are provided (beat and chord from MIDI/audio using MIR methods).
• Two-team quality control yields high-quality, style-consistent piano arrangements suitable for arrangement tasks.
• Baseline transformer experiments on a MIDI-like event representation show train loss 2.08978 (Train) and 2.38122 (Test) with accuracies 0.62021 (Train) and 0.54529 (Test).
• Conditional piano-arrangement generation conditioned on melody demonstrates consistency in harmonic relations and rhythm, illustrating dataset utility for Task 1 and Task 2."
• The dataset supports both arrangement generation and cross-modal generation (e.g., audio-to-piano reduction) and serves as a richer resource than prior datasets for arrangement-focused research.