A dataset for combinatorial music generation, presented at the NeurIPS 2022 Datasets and Benchmarks Track. It contains 11,144 composer-made MIDI samples, each labeled with 12 types of metadata.

Key contributions

  • 01

    Combinatorial music generation

    The paper introduces this task: a model generates short note sequences from metadata, and the sequences are combined into a complete song.

  • 02

    Composer-made dataset

    Professional composers wrote every sample by hand, following written guidelines to keep the data consistent.

  • 03

    12 types of metadata

    These include BPM, key, instrument, and nine others. Two are unique to ComMU: track role and chord progression with extended chord qualities such as sus4 and maj7.

How it works

  • Stage 1

    Generate note sequences

    A model generates a note sequence for each set of metadata.

  • Stage 2

    Combine into a song

    For now, composers arrange the generated note sequences across tracks and time to complete a song. Automating this stage is future work.

Metadata

BPM
80
Key
A minor
Time signature
4/4
Number of measures
4
Genre
cinematic
Rhythm
standard
Track role
main melody
Pitch range
mid
Instrument
violin
Min/max velocity
2, 127
Chord progression
C-Am-F-G

Note sequence

aSLwZ5Ts

Combination

t1t2t3
Main melody14…
Sub melody…
Accompaniment222…
Bass…
Riff33…
Pad5…

Each number refers to a note sequence generated in Stage 1

Dataset

Over six months, 14 professional composers worked in two teams. Team A analyzed reference music and wrote composition guidelines, and Team B composed samples following them.

Team A

  1. 01Collect reference music by genre and mood
  2. 02Analyze the reference music
  3. 03Write composition guidelines with metadata

Team B

  1. 04Compose MIDI samples following the guidelines

ComMU dataset

11,144

MIDI samples with corresponding metadata

  • Track role

    Each sample is labeled with its part in an arrangement: main melody, sub melody, accompaniment, bass, pad, or riff. With instrument labels alone, a model knows to use piano and guitar but not what each one plays. Adding track role lets it assign accompaniment to the piano and melody to the guitar.

  • Chord progression

    Each sample comes with the sequence of chords it uses. Besides major, minor, diminished, augmented, and dominant, ComMU includes sus4, maj7, half-diminished, and min7. These extended chord qualities let the model learn more melodies for the same progression, improving the harmony and diversity of the music it generates.

  • Genre

    The dataset covers two genres often used in background music. New age is melody-driven, built on keyboards and small acoustic instruments. Cinematic is orchestral, with string ensembles carrying the melody and accompaniment.

Results

  • Fidelity

    In a survey, 40 anonymous composers compared generated and human-composed samples made from the same metadata and chose the one they preferred. Generated samples won about 42% of the time.

    Human-composed

    ≈58%

    Generated

    ≈42%

    0255075100%

    Values read from the paper's chart

  • Controllability and diversity

    Controllability is the proportion of generated notes that follow the pitch range, velocity range, and key or chord specified in the metadata. Diversity is the average pairwise distance between note sequences generated from the same metadata.

    Pitch control
    0.84
    Velocity control
    0.91
    Harmony control
    0.99
    Diversity
    0.32

    Sampled with top-k 32, temperature 0.95

Seoul Office, South Korea 302, 10 Nambusunhwan-ro 333-gil, Seocho-gu, Seoul, 06725

598-81-00957, Wongil Huh

contact@pozalabs.com

Copyright © Pozalabs inc. All rights reserved