Google created an artificial intelligence capable of making songs based on text descriptions. The project, documented in a Jan. 26 post, is called MusicLM and is “a high-fidelity music generator model from text descriptions,” as the company explains.
- What is Bubble Tea, the drink celebrated by the Google Doodle
- Google app receives more visual tweaks to the Material You pattern
MusicLM can create tracks in 24 kHz in a wide variety of styles, following the characteristics presented by the user. Whether it’s soundtracks from classic arcade games, a guitar solo, jazz or electronic music, the possibilities with MusicLM are virtually endless (within the scope of AI, obviously).
In addition to short songs, Google’s AI can also create longer tracks, lasting up to five minutes. By describing stories and narrations, the project is also able to create soundtracks to accompany the tale.
Music AI training
According to the MusicLM documentation, the model was fed about 280,000 hours of music content to learn how to generate coherent (and significantly complex) melodies for each context.
As a result, AI creations can’t be as original, and savvy music users can find similarities to specific tracks and artists fairly easily. Even so, MusicLM doesn’t actually plagiarize the sounds, but composes based on existing tracks.
In addition to following text descriptions, MusicLM can also create tracks based on images or generate melodies from a given instrument. The AI is refined enough even to create songs that hark back to a specific era.
Some of MusicLM’s creations have been showcased on GitHub. Each of the demos manages to follow the text descriptions precisely, while not being entirely new in style.
Furthermore, it is not possible to know whether AI can deliver results with such quality in all cases. Since Google only highlighted a few tracks, it’s possible that the company only chose the best creations. Vocals, for example, are not a strong point for MusicLM and some samples present distortions at certain moments (a consequence of the work being done by AI, not humans).
As with the issues raised by ChatGPT and authorship over text projects, MusicLM also has the potential for controversy: Google researchers claim that around 1% of AI results were nearly identical copies of the source material.
“We recognize the risk of potential misappropriation of creative content associated with the use case,” explained the AI co-authors. “We strongly emphasize the need for more future work to address these risks associated with generating music.”