google-research.github.io
MusicLM is a model that generates high-fidelity music from text descriptions. It frames conditional music generation as hierarchical sequence-to-sequence modeling and produces audio at 24 kHz consistent over several minutes. It can also be conditioned on text and a melody, transforming whistled or hummed melodies according to a text caption. The project released MusicCaps, a dataset of 5.5k music-text pairs with descriptions by human experts.