Don’t shoot the messenger, please. Google’s MusicLM claims to have finally reached the peak of all of the music-generating AI systems previously released. MusicLM can take a prompt from a variety of text prompts, including simply naming a genre, to the description of a classic painting, to a vibe—such as “beach in the caribbeans (sic),” to much longer, more descriptive prompts.
An example of a longer prompt, or “Audio generation from rich text captions,” as Google calls it, is below and sounds like this.
Epic soundtrack using orchestral instruments. The piece builds tension, creates a sense of urgency. An a cappella chorus sing in unison, it creates a sense of power and strength.
Like it or not, AI is here to stay, and it continues to creep closer to convincingly spitting out music, text, and art based on hundreds of thousands of samples, teaching this artificial life to imitate writing, visual art, and music that gets increasingly better than previous attempts.
With Microsoft sinking $10B into OpenAI’s ChatGPT, this technology isn’t going away any time soon. With such big players as Google and Microsoft coming to the table with their deep pockets, ready to play, we will continue to see this technology evolve.
Borrowing from existing sources of music, Google’s programmers have fed MusicLM enough data that the technology gets eerily close on some of the types of music or responding to a prompt. It’s enough to freak professional creatives out.
Could MusicLM replace, say, songwriters, and musicians? Google pinkie swears that’s not in the plans for the immediate future. However, as Google has fed MusicLM actual, existing songs to teach it how to create audio excerpts, they report that around 1% of the “music” generated is a direct copy of a real song.
Thus, Google claims they are not releasing the system to the public. Regardless, a quick search on Twitter unveils many passionately rejecting the tool.
The technology hits closer on some types of music that are electronic to begin with. See for yourself on Google’s MusicLM Github, where they have released a plethora of examples, showing interested parties the prompt and the music. Compare the accuracy of “berlin 90s house” with the way off-base “blues” for a feel at how well the AI does across the board. It may be more accurate than previous attempts, and some of it is convincing. It’s a version of music.
Among the dozens of MusicLM audio files released by Google, there aren’t any true bangers there. Yet.
Some hit closer than others, clearly. Because appreciating music is personal and subjective, some may think MusicLM’s versions would do in a pinch, just as some are marveling at how well ChatGPT can write a poem or social media caption. It will continue to learn and improve as it gets fed more and more authentic examples of “real” music and responds to more prompts.
The interpretation of classic paintings, based on descriptions of the paintings, will be a source of contention. After all, much like watching a movie based on one’s favorite novel, it all comes down to how well done the presented interpretation is and how closely it mirrors the images and feelings the original novel or painting conjured up in our own minds and hearts.
Does the MusicLM audio interpretation of Picasso’s bleak “Guernica” depict, for example, “…the suffering wrought by violence and chaos”? You be the judge—but be sure to listen all the way to the end of the 30 second “Guernica” clip for the voices before settling on your decision.
It will take time to see if, when, and how Google releases MusicLM for general use by the public.
Like Pac-Man voraciously munching those dots and ghosts, AI systems need to be constantly fed data in order to learn more and improve. 100% of that data comes from real, live human songwriters and musicians’ intellectual property that lives on the internet. Can Google legitimately claim, then, that a mere 1% of it is directly lifted from existing content?
That’s their story, and they’re sticking to it for now. The powers that be at Google are wise to cool their jets while waiting to see what happens with a lawsuit against OpenAI, Microsoft, and Github accusing ChatGPT of intellectual property violations and a similar lawsuit brought by artists against DeviantArt, Stability AI, MidJourney (three text-to-art generators that use Stable Diffusion, which trains these tools by feeding it millions of pieces of artwork that exist online).
Joleen Jernigan is an ever-curious writer, grammar nerd, and social media strategist with a background in training, education, and educational publishing. A native Texan, Joleen has traveled extensively, worked in six countries, and holds an MA in Teaching English as a Second Language. She lives in Austin and constantly seeks out the best the city has to offer.
Pingback: Is Google’s music AI generating the sounds of the future? - Namecheap Blog