Emotion Labels Dataset Prototype for AI Music Generation
(2019)
Research Intern @ DeepMusic Co. Ltd.
Embedded Files
To prototype a new dataset for affective AI music generation from emotion label prompts, this project proposes a dictionary of many-to-many mapping rules that directly correlate emotion labels with musical elements based on case studies of pop music and film music, setting a standard for future addition of more complicated rules and models at the micro and macro level.
KEYWORDS:
AI Music Generation; Feature Engineering for Dataset; Emotion Classification
Robert Plutchik's Wheel of Emotions. By Machine Elf 1735 - Own work, Public Domain, https://commons.wikimedia.org/w/index.php?curid=13285286
Motivation and Inspiration
Motivation and Inspiration
As music classification becomes more diverse, users rely increasingly on emotion labels to search for music recommendation on streaming applications and stock music platforms. Researchers in music psychology are also expected to benefit from a more nuanced dataset with a wider variety of audio samples, detailed emotion labels, and controllable musical variables.
This project uses a multi-level emotion label system where the primary labels are inspired by Robert Plutchik's three-dimensional model; complex emotion labels are stemmed from one or more primary labels, forming a system of many-to-many mapping.
Mapping Emotion Labels to Musical Elements
Mapping Emotion Labels to Musical Elements
The mapping rules directly correlate emotion labels with musical elements, which are further translated to data representations readable to the AI music generation models.
Emotion labels are collected from major streaming services (e.g., Netease Music, QQ Music) and stock music platforms (e.g., Vfine Music, Splice). Music examples for case study are sourced from chart-topping popular music across multiple genres (e.g., pop, rock, EDM) as well as famous film music (orchestral and synthesized).
Musical elements considered include:
Tempo (BPM)
Chord progressions (starting with triads and 7th chords)
Chord dynamics (e.g., arpeggios, tutti )
Instrumentation (e.g., piano, synth)
Melody progression (e.g., harmonic/inharmonic intervals)
Performance dynamics (pp to ff )
An example case study may look like follows:
Music: stock music, Halloween-themed
Labels: happy, cheerful, magical, child-like, festive, spooky
Tempo: 72-108 BPM
Instrumentation: piano, pitched percussion (glockenspiel, marimba), synth
Chord: Maj 7th, arpeggios for perc. instruments
Others: (include specific synth loops as reference)
An excerpt of the worksheet (in Chinese) of summary of mapping rules.
Page updated
Google Sites
Report abuse