AI Briefing[NeurIPS 2022] Symbolic Music Motif Generation - LG AI Research BLOG
Key point
LG AI Research has released research on a model that effectively generates Loops, the core unit of music.
Details
To generate the Motif, the first step of music creation, we introduce research that focuses on generating Loops of 4 to 8 measures rather than an entire song. Data is processed using a Binary Representation that quantizes MIDI data into sixteenth-note units.
To reflect the repetitive patterns of music, a One-class Loop Detector was designed. Using anomaly detection principles based on Deep SVDD, it learns the structural rules of music and demonstrates performance in determining whether something is a loop.
The generative model adopts an approach where VQ-VAE compresses the loop set, and the compressed Discrete Latent Code is modeled with an LSTM. This allows the loop, the core skeleton of music, to be represented in a finite space.
Performance evaluation results showed superior performance in terms of sample fidelity and diversity compared to Music Transformer or MuseGAN, and positive results were also obtained in human listening tests.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.