Google Creates AI-Powered Music Maker, But it Won’t Be Released

If you thought OpenAI’s ChatGPT is impressive, wait till you see what Google has been working on — if the company lets you, that is.

Among Google’s many endeavours in artificial intelligence (AI) and machine learning is MusicLM, a model the company created to generate original music from rich text prompts (via TechCrunch).

Google researchers detailed MusicLM in a recently published whitepaper. The model was trained on a dataset comprising 280,000 hours of music to learn to compose coherent songs with — as the researchers put it — “significant complexity,” from text prompts like “a calming violin melody backed by a distorted guitar riff.”

MusicLM manages to do just that and more. It is arguably the first AI model that produces high-fidelity 24 kHz music that is complex in composition and actually sounds like what a human artist might create.

MusicLM can even work with long prompts, accurately capturing instrumental riffs, melodies, and moods in its music. Admittedly, the model is nowhere near as adept with vocals as it is with instrumentals. While it is able to create vocals, they sound warped at best and are nowhere near as coherent as the music it synthesizes.

Even so, MusicLM handily surpasses competitors like Dance Diffusion, OpenAI’s Jukebox, and Google’s own AudioML. Check out the Google Research whitepaper on MusicLM, which includes dozens of AI-generated music samples.

MusicLM can even generate music from prompts that are a combination of text and image, and accurately capture what an instrument would sound like in a certain genre. Users can even specify the level of music experience they want the generated tune to reflect.

However, the AI-powered music maker won’t see the light of day in its current form. The researchers who co-authored the whitepaper noted that they have no intention to release MusicLM because of the “potential misappropriation” of copyrighted content.

During an experiment, the researchers found that about 1% of the music generated by MusicLM was an exact duplication of the songs it was trained on.

“We acknowledge the risk of potential misappropriation of creative content associated to the use case,” the paper reads. “We strongly emphasize the need for more future work in tackling these risks associated to music generation.”

Click here to listen to examples of MusicLM-generated music, and tell us what you think about them in the comments below.

P.S. - Like our news? Support the site with a coffee/beer. Or shop with our Amazon link. We use affiliate links when possible--thank you for supporting independent media.