Producing music from a description or a fragment. Quality is high enough for background and production use, and the field is dominated by the rights question, since training on recorded music without licence is the subject of active litigation and of licensing deals being struck in parallel.
Technical view
Approaches include token-based modelling over neural audio codecs and diffusion in latent audio space, conditioned on text, melody or reference audio. Technical challenges are long-range structure, since coherent musical form spans minutes, and controllability of specific instrumentation and arrangement. Commercially the binding constraints are training data licensing, output similarity to protected works, and platform policies on distribution of generated tracks.
Related
Part of AI Terminology Index, section: Modalities and Applications.