Andrea Agostinelli, Timo I. Denk, Zalán Borsos
Jesse Engel, Mauro Verzetti, Antoine Caillon, Qingqing Huang, Aren Jansen, Adam Roberts, Marco Tagliasacchi, Matt Sharifi, Neil Zeghidour, Christian Frank
2023 · arXiv
We introduce MusicLM, a model generating high-fidelity music from text descriptions such as "a calming violin melody backed by a distorted guitar riff". MusicLM casts the process of conditional music generation as a hierarchical sequence-to-sequence modeling task, and it generates music at 24 kHz that remains consistent over several minutes. Our experiments show that MusicLM outperforms previous systems both in audio quality and adherence to the text description. Moreover, we demonstrate that MusicLM can be conditioned on both text and a melody in that it can transform whistled and hummed melodies according 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.
Metadata: arXiv. Abstracts retain their original authors’ rights.
Original HN comments, with labels from simple text rules. A label describes the whole comment and may miss context or irony. Read the discussion to make up your own mind. About the labels
3 selected excerpts · 3 total mentions
"Computer Music" is a very broad term (no surprises here) so, like many here, I can point out topics that are not covered. In particular, computer music (aka algorithmic) composition [1], or very recent AI techniques like the Google seq2seq example at [2], or the (unpublished, but probably a form of generative adversarial networks) techniques used by SunoAI and Udio. [1]. https://ccrma.stanford.edu/~blackrse/algorithm.html [2]. https://arxiv.org/pdf/2301.11325
>"Nice, but somehow it feels like I've heard these songs before. Could that be possible?" According to the paper[0] "We found that only a tiny fraction of examples was memorized exactly, while for 1% of the examples we could identify an approximate match." FWIW I'm pretty sure I heard a fragment of Kalinka (a folk tune in the public domain) in one of the samples I generated. [0] https://arxiv.org/pdf/2301.11325.pdf
Link to paper: https://arxiv.org/abs/2301.11325 (the one on the site seems broken) Even with the knowledge that something like this is inevitabely coming, it's pretty crazy to see the examples. The fact it can create fitting music from painting descriptions (including ones such as or Guernica) is unexpected. And it seems to generate almost naturalistic vocals with reasonable-sounding lyrics, with no embedded structure for lyrics or anything. The paper itself is pretty cool. While GPT-3 was just "GPT-2, but bigger", this seems to have required some clever combination of models. Wonder if similar ideas will be needed for longer-form text or video generation.
More from the same topics.
Dan Tong, David J. Farnham, Lei Duan +4
Abstract If future net-zero emissions energy systems rely heavily on solar and wind resources, spatial and temporal mismatches between resource availability and electricity demand may challenge system reliability.
Steven M. Paul, Daniel S. Mytelka, Christopher T. Dunwiddie +4