Abstract

We introduce MusicLM, a model generating high-fidelity music from textdescriptions such as "a calming violin melody backed by a distorted guitarriff". MusicLM casts the process of conditional music generation as ahierarchical sequence-to-sequence modeling task, and it generates music at 24kHz that remains consistent over several minutes. Our experiments show thatMusicLM outperforms previous systems both in audio quality and adherence to thetext description. Moreover, we demonstrate that MusicLM can be conditioned onboth text and a melody in that it can transform whistled and hummed melodiesaccording to the style described in a text caption. To support future research,we publicly release MusicCaps, a dataset composed of 5.5k music-text pairs,with rich text descriptions provided by human experts.