Will artificial intelligence be able to replace human creation, also in the field of music? (Photo: Unsplash)
"A rhythmic and upbeat arcade game soundtrack, with a catchy electric guitar riff"
"A magical jazz song with an unforgettable sax solo and lead singer"
"90s Berlin techno with low bass and a strong kick."
All this and much more you can create in MusicLM, Google's artificial music generator, which is currently in accelerated development stages.
With a dataset of 280,000 hours of music, this system was specifically trained to learn to create coherent songs from descriptions with "significant complexity," according to the creators.
His music, surprisingly, sounds like something a human artist might have composed, though it's not necessarily innovative or musically cohesive.
Given that there are no real musicians or players in the loop, it's hard to overstate the quality of the samples.
Even when fed with long and rambling descriptions, MusicLM manages to capture nuances such as instrumental riffs, melodies and moods.
In this way, musicians, as well as advertisers, brands, or anyone who makes commercial use of music, will be able to produce sound-based messages and plan them precisely to motivate one or another target audience to the desired action.
In the company's announcement, Google's researchers demonstrated the diverse capabilities offered by the system:
Identifying and performing existing melodies by humming, whistling, singing or playing
System guidance through a combination of image and text
Creating audio that is "played" by a certain musical instrument in a certain genre, and defining the experience level of the "musicians"
Creating music inspired by places, historical periods, specific artists, styles, or a desired effect on the audience such as "motivation for fitness training" or "opening the heart and spiritual ascension".
Enter several descriptions in sequence into a "story" or melodic narrative several minutes long, for example "time to wake up, time to meditate, time to run, time to give 100%".
Marketing applications of the music generators
The last option makes MusicLM particularly convenient for creating a soundtrack for a movie, or in the more likely case, YouTube videos, which suggests another genius step by Google in its strategic game against its competitors in the social field - Tiktok, Meta, etc.
When the faces of all the platforms are directed towards monetization from advertising video ads, creating music in this way may be the solution for using musical pieces in advertisements, which are both original, approved for use, and quick to prepare.
But MusicLM is not the first artificial intelligence system that generates music.
Previous attempts include Riffusion, which composes music based on visual imaging, Google's Dance Diffusion and AudioML, and OpenAI's Jukebox.
But due to technical limitations and limited training data, none of them were able to produce particularly complex songs.
An academic article about the technology claims that this is probably the first system in the world to be able to do this, but it is more likely to assume that, at least in the near future, most applications of artificial music generators will be in the field of digital marketing and advertising and less in the field of artistic creation.
at a special sale price
The revolution of facial treatments at home: artificial intelligence and radio waves
Served on behalf of B Cure Laser
Quantity instead of quality
Just as ChatGPT relies on cultural biases and Midgerani doesn't know how to draw hands, an unavoidable side effect of the training process makes some of the samples in MusicML sound distorted.
Although she can produce vocals, including choral harmonies, most of the "lyrics" range from barely English to pure gibberish, sung by synthesized voices that sound like amalgamations of several artists together.
As of today, OpenAI seems to be a few steps ahead of Google, with more cohesive musical pieces and an experimental version already available to developers, though neither has launched an official product yet.
It is very possible that this announcement by Google about the technology, before a launch date has been set, is an attempt to show a presence in a market that might otherwise grow very quickly without it.
Indeed, in view of the threat of ChatGPT, Google has already announced an emergency meeting of the company that will speed up the development processes of several future products, and this in order to remain a leader in its field.
At the same time, we have already seen in the past how Google eliminates products after investing millions, because they were not profitable enough or the market was not mature, therefore, until we see a product available, everything should be taken with a limited warranty.
In fact, the opposite process is currently happening, and for the first time in its history, Google is actually lagging behind the competition.
Therefore, it is no wonder that the quality of the music, as of now, is not a top priority.
Copyright and Deepfake
But the biggest obstacle of this technology is still ahead.
After Jukebox's release, critics questioned whether training AI models on copyrighted musical material constituted fair use.
Similar concerns have been raised about the training data used in artificial intelligence systems to generate images, code and text, which are often collected from the web without the creators' knowledge.
Assuming MusicLM or a system like it becomes available one day, it seems inevitable that major legal issues will come to the fore, even if the systems are marketed as an aid to artists rather than an artificial replacement for them.
Google's researchers anticipate a cure for the blow and openly point out the many ethical challenges posed by the system, including the tendency to incorporate musical material from the training data into new compositions.
In an experiment conducted by the company, it was discovered that about 1% of the music the system created was copied directly from the songs it was trained on.
Since these songs are copyrighted, this threshold is likely high enough to prevent Google from launching MusicLM in its current state.
"We acknowledge the risk of potential exploitation of creative content related to the use case," the paper's authors wrote.
"We strongly emphasize the need for further future work addressing risks associated with music making."
Google of course speaks based on experience.
In 2020, Jay-Z's record label filed a copyright lawsuit against YouTube channel Vocal Synthesis, which specializes in music deepfakes, which used artificial intelligence to create Jay-Z covers of songs like Billy G's "We Didn't Start the Fire" 'Al.
After initially removing the videos, YouTube reinstated them, claiming that the removal requests were "incomplete."
This case demonstrates how the whole issue of deepfake music still stands on very shaky legal ground.
A document authored by Eric Sunrey, now a legal intern at the Association of Music Publishers, claims that artificial music generators like MusicLM infringe on musicians' copyrights by creating "coherent audio tapestry of the compositions they consume in their practice, thereby violating United States copyright law."
Experts speculate that music created by an AI system would be considered a derivative work, in which case only the original elements would be protected by copyright, although it is unclear what might be considered "original" in such music.
What is certain is that the commercial use of artificial music, other than for commentary or parody purposes, is entering uncharted waters, and the courts will have to rule on each case individually until there is legislation on the matter.
A number of lawsuits already making their way to the courts will likely have a decisive impact on the future of technology, and will determine when and if at all we will get to use these tools in practice.
Finally, a playlist created entirely using OpenAI's Jukebox
OpenAI · Jukebox samples: Novel lyrics
Marketing and digital
in the headlines