A 132-page monograph, according to the submitter, on neural methods for music audio generation, aimed at people who know neural-network training but are new to music.
Covers: codecs, autoencoders, token models, diffusion and flow matching, sampling speed-ups and preference optimization, in one notation with worked examples.
Applications: instrumental clips, full songs with vocals, editing and live performance, plus comparison tables of published systems.
Evaluation and law: reading metrics and learned judges, listening tests, data provenance, memorization, attribution and selected legal disputes.
Open problems: the submitter lists long-range song structure, evaluation, memorization, language coverage and interpretability.
Sign in to comment.
TL;DR of the paper, generated automatically.
A 132-page monograph, according to the submitter, on neural methods for music audio generation, aimed at people who know neural-network training but are new to music.