symbolic_music.md

link

• ISMIR23-intro1, ISMIR23-intro2

Representation

Piano-roll like

2-dimensional matrix, where one dimension represents time and the other represents pitch

Event-level

• MIDI-like(Music transformer)
  • including note-on, note-off, velocity, pitch, etc.
• REMI(Pop Music Transformer)
  • improves the MIDI-like representation by using duration, bar, chord, and tempo
• Structured MIDI
  • use time-shift tokens instead of note-on/off or duration tokens

Note-level

• MuMIDI(PopMAG), CP(Compound word transformer), OctupleMIDI(MusicBERT), Multitrack Music Transformer
  • compress the attributes of a note, including pitch, duration, and velocity into one symbol

Generation metrics (General)

Negative log-likelihood (NLL)

• smaller values are better
• e.g., Music Transformer

perplexity

• smaller values are better
• e.g., Museformer

empty bars (EB)

• the ratio of empty bars
• values closer to the original data are better
• e.g., MuseGAN

used pitch classes (UPC)

• the number of used pitch classes per bar
• values closer to the original data are better
• e.g., MuseGAN

qualified notes (QN)

• the ratio of qualified notes that are no shorter than a time step(i.e. a 32th note)
• values closer to the original data are better
• e.g., MuseGAN

Drum Pattern (DP)

• the ratio of notes in 8- or 16-beat patterns
• values closer to the original data are better
• e.g., MuseGAN

Tonal Distance (TD)

• the harmonicity between a pair of tracks
• smaller values are better
• e.g., MuseGAN

Beat/Downbeat STD

• convert the generated symbolic music into audio music and evaluate by beat tracking model
• values closer to the original data are better
• e.g., Pop music transformer

Note F1

• match if generated onset, measure, pitch, track match exactly those of original
• higher values are better
• e.g., MT3, Composer Assistant

Pitch class histogram entropy difference

• Drums are ignored
• lower values are better
• e.g., Composer Assistant, XLNet

Groove similarity

• higher values are better

MIREX-like prediction test(≒PPL, NLL)

• choose the correct answer from 4 choices by calculating the average probability of generating the event
• e.g., Jazz transformer, XLNet Piano Infilling

Human Eval

• number of wins

  • e.g., Music Transformer

• harmonious/rhythmic/musically structured/coherent/overall rating

  • e.g., MuseGAN

• musicality/short-term structure/long-term structure/overall/overall score/preference

  • e.g., Museformer

• structure/richness/pleasure/overall

  • e.g., MELON

• coherence/richness/arrangement/overall

  • e.g., MMT

• average rank, p-value

  • e.g., composer's assistant

• distinguish pro and non-pro, p-value

  • e.g., Pop Music Transformer

Similarity Error (SE)

• the error between the similarity distribution of original data and generated music
• smaller values are better
• e.g., Museformer

Inference speed

• notes per second
• e.g., MMT

Generation Analysis

similarity distribution

• to analyze the repetition
• e.g., Museformer

Attention gain

• e.g., MMT