Meldy: a music generator

Matteo Bernardini, Yilin Zhu
{10743181,10702368}@mail.polimi.it

ACTAM & CMRM Project 2019-2020

Overview of the Project

Project Structure

Topics

  • Computer Music
  • Theoretical C.S.
  • Computational Creativity
 

Resources

  • music21: computer-aided musicology toolkit
  • p5.js: web visualization
  • OSMD: MusicXML rendering
  • webpack: development

1) User Input

  • 2D picker: strength of a mood
    • implemented using p5.js

 

  • Dimensional approach from MER
    • x-axis: valence, in range [0, 1]
    • y-axis: arousal, in range [0, 1]

Choose a mood as input

2) Melody Generation

2.1 Mapping from mood to music features

  • valence & arousal
  • higher valence -> brighter mode
  • higher arousal -> higher pitch
  • higher arousal -> faster tempo

2) Melody Generation

2.2 Time signature

 
  •   different moods -> different rhythmic patterns
    • 4/4 : stability
    • 3/4 : danceability
    • odd time signatures (e.g. 7/8): tension and instabilities
       

We decided to ignore this aspect and simply stick on a 4/4, in order to have a better focus on the other main features and simplify the approach for durations generation.

2) Melody Generation

2.3 Key signature

  • mode is decided before, we need a root note
  • our approach: pick a random root within the possible 12 pitches of the cromatic scale
  • first problem: Not all possible combinations of root names and mode are theoretically valid
    • usual names -> C C# D Eb E F F# G Ab A Bb B
  • second problem: these are usual names only for ionian mode
    • pick random "base root" in ionian -> rotate to mode (e.g. base: C#, mode: lydian -> key: F# lydian)

2) Melody Generation

2.4 Melodic Sequence

  • Base model: L-system (Formal Grammar)
    $$ G = \{V, \omega, P\} $$
  • \(P\) and \(\omega\) are stochastic
  • Extended model: different rulesets and a selection parameter
    $$ G = \{V,\omega,R,p\} \qquad R = \{ (P_1,s_1), (P_2,s_2),\cdots\} \qquad p \in \mathbb{R} $$
  • Two separate grammars for generating:
    • Sequence of pitches
    • Sequence of durations

2) Melody Generation

2.4 Melodic Sequence

  • \(p\) is used to select which \(P_i\) to use
  • \(P_i\) is used for the grammar if \(|p-s_i|\) is the minimum difference among the defined rulesets

 

  • This model introduces further variability over a basic L-system
G = \{V,\omega,R,p\} \qquad R = \{ (P_1,s_1), (P_2,s_2),\cdots\} \qquad p \in \mathbb{R}

2) Melody Generation

2.5 Pitches

  • mood from input -> Key signature
    • use relative degrees instead of actual pitches
    • we can do the transposition later
  • valence influences the melody (i.e. p=valence)
    • positive -> uplifting
    • negative -> downwards
  • vary between two rulesets:
    • higher valence (s=1)
      higher probabilities to be uplifting, create consonant melodic intervals.​
    • lower valence (s=0)
      downwards, dissonant melodic intervals 

2) Melody Generation

2.6 Durations

  • simplify the task:
    • stick to 4/4
    • define durations with "1" = "quarter note"
  • vary between 3 rulesets (p=arousal):
    • ​higher arousal (s=1)
      add faster notes (8ths and 16ths) and syncopations
    • middle arousal (s=0.5)
      remove faster notes and syncopations
    • low arousal (s=0)
      simple and straight rhythms

2) Melody Generation

2.7 Output format

  • MusicXML: aimed to music representation, better suited than MIDI
  • Natively supported by music21

3) Resources

  • First attempt: music21j
    • not mature enough, several bugs
       
  • Conclusion: music21 (python version)
    • back-end needed for this step

3.1 Musicology Tools: Music21

3) Resources

3.2 Development Pipeline

4) Music Rendering

  • User clicks the Impress Me button → view switch
  • Rendering MusicXML to Music Notation
    • OpenSheetMusicDisplay: TypeScript Library
    • output as SVG, rendered natively by browser

4.1 Show Sheet Music

4) Music Rendering

  • Play → listen to the generated music
  • Download save MusicXML (open with Finale, Sybelius, ...)
  • Restart → go back to the mood selection view

4.2 Play and listen to it

Future Works

  • Enhance grammar model accuracy
    * try to apply some evaluation or reward function
  • From single melodic paragraph to multi-paragraphs
  • Introduce different musical forms (e.g. sonata)
  • Other music elements (e.g. different time signature)

How can we improve this project

Reference & Links

Thanks for your attention!