AIs are on the rise. After generating an image of MidJourney or the conversation held by Chat GPT, Google stands out with a new tool capable of generating music from… text. If the idea is not new, since AudioLM (Google) or JukeBox (OpenAI) are already capable of it, MusicLM, it is its small name, has like ChatGPT a huge database ( we are talking about 280.000 hours of music) in order to be as relevant as possible. And it works !
A Github page offers, if you don't use the tool, to listen to some examples. The researchers behind MusicLM also published their study on the subject on Arxiv.org, in a fifteen-page PDF, which explains the concept:
"We present MusicLM, an AI model capable of generating high-fidelity music from text descriptions such as 'a soothing violin melody backed by a distorted guitar riff'. MusicLM treats conditional music generation as a sequential hierarchical modeling task and generates 24kHz music that remains consistent over several minutes. Our experiences show that MusicLM outperforms previous systems both in audio quality and in compliance with text descriptions. Additionally, we demonstrate that MusicLM can be conditioned by both text and melody, as it can transform whistled and hummed melodies to the style described in a text caption. To support future research, we are releasing MusicCaps, a dataset consisting of 5,5k music-text pairs, with rich text descriptions provided by human experts. »
The future of the artistic seems more and more to be written in 2 letters. Although AI will probably never replace flesh-and-blood artists, it is a great help and an endless source of inspiration.
