A few days ago a song titled Permission is hereby granted, whose text recites in full the MIT license for the distribution of software has gone viral, at least in the geekiest corners of the Internet. The piece is a poignant ballad, in which a female voice reminiscent of Imogen Heap emphatically recites unlikely phrases in legalese technological. A funny meme, especially for the contrast between the musical style and the technical and intricate text. However, the singer whose voice is heard does not exist and no musician has ever composed the melody and accompaniment, nor played any instrument.
The song was created with the latest version (v3) of sun, a service that allows you to create entire songs, up to two minutes long, in any style and of any genre using a proprietary generative artificial intelligence model. The software was created by the startup of the same name founded in Cambridge, in the US state of Massachusetts, by Michael Shulman, Georg Kucsko, Martin Camacho and Keenan Freyberg. The 4 all worked for Kensho, another AI startup, and boast previous professional experience at companies like Meta and TikTok.
Publicly unveiled in December 2023, Suno (a term meaning “to listen” in Hindi) made headlines for a lightning-fast partnership with Microsoft. The Redmond company immediately signed an agreement with the startup to integrate the music generation service into Copilot. It is only with the release of version 3 of the model, however, that Suno’s popularity is exploding. Now the service allows you to create songs up to two minutes long which can then be made even longer, and for now it’s all free (with some limits).
Artificial intelligence
The New York Times sues OpenAI and Microsoft for copyright infringement
by Emanuele Capone
How Suno works
To use the service you must going up Suno.com e fare login with a Google, Microsoft or Mastodon account. Once connected, you receive 50 credits which are renewed every twenty-four hours. Each generation costs 10 credits and produces two different versions of the same piece. To create a song, all you need is a few lines in which you specify the genre, the rhythm you want to obtain, and the type of song (such as ballad, disco hit and so on). Alternatively, AI can do everything on its own: in that case, however, a completely random piece is generated. The interpretation of prompts and the creation of texts is done thanks to Suno’s integration with ChatGPT.
Suno’s strong point is the Custom mode, which allows you not only to specify the genre and type of song, but also to provide a complete text in which you can indicate verses and chorus with the Verse Chorus indications. If you also want to add any choirs, simply indicate the words from make virtual vocalists sing in round brackets. It is in this way that the aforementioned viral song with the lyrics of the MIT license was created. The system not only works in English but also accepts and correctly generates texts in other languages, including Italian.
The songs generated by Suno are not exactly chart hits, but with some forethought and trying to play with indications such as Chorus and Verse you get catchy pieces. The result, judging by the most popular songs in Suno’s public archive, do not have high-level artistic aspirations, but are rather musical memes that aim to make you laugh or surprise. There are of course also various declinations of the internet theme par excellence, that is the kittens.
Our song, Lights in the circuit
We have too tried to write a song with Suno, and we did it (for consistency) by asking ChatGPT for help in writing the text. We ordered the OpenAI model to create verses that spoke about the future of artificial intelligence and the lights and shadows of a technology that promises to change the world. As you may have guessed, we were not aiming for originality in any way. We then took the text, added the directions for the division into verses and refrain and finally fed it all to Suno.
The result is a song titled (from ChatGPT) Lights in the circuit. In the Suno prompt we specified that the style should be that of an Italian pop ballad, with a medium speed and a piano base. The result seems like a filler song taken from a generic Italian pop album from the early 2000s: it won’t win Sanremo but it shouldn’t be thrown away either. ChatGPT, when asked directly, suggests that the names of the authors of this piece could be Alisei, Vetro and Dumbledore. And therefore, without further ado: by Alisei-Vetro-Silente, Lights in the circuitMaestro Suno conducts the orchestra.
The unsolved problem of copyright
Boutades and memes aside, Suno AI is an amazing technology. And the first generative artificial intelligence model capable of creating songs that have a very high coherence between melody and lyrics, all while maintaining an amazing adherence to the style specified in the prompt.
But There’s a problem, and it’s a big one: no one knows how the model underlying Suno was trained, and above all whether copyright-protected music libraries for which the startup had no license were used for the training. Or rather: i founders they know it about Suno, and probably so do they investors that millions have already covered them, but they don’t want to say it.
May Suno have learned to compose music studying more or less famous songs is almost incontrovertible. Composer and AI programmer Ed Newton-Rex, former Vice President of the audio division of Stability AI and now CEO of the startup Fairly Trained, has conducted an in-depth analysis of the results that can be obtained with Suno. In some cases he managed to force the model to spit out pieces that feature famous chords, arrangements and melodies, overcoming the filters imposed on the prompt by the company which prevent the names of many famous artists from being used in the prompts.
The similarities are undoubted and if the generated pieces were written by human musicians they would most likely be branded as cases of plagiarism. According to Newton-Rex, the fact that the founders do not want to answer the direct question about the copyright of the data used to train the model does not necessarily indicate that there were no licensing agreements. However in a quote collected from Rolling Stone, one of the early investors in the startup suggests that there were no deals with record companies at the time the company was founded. The investor also reportedly expressed awareness that the majors could soon sue Suno. All indications that support the hypothesis of a model training occurred without any concession on rights.
These uncertainties make it difficult to even answer another fundamental question: Who owns the songs generated by Suno? In a section of the site dedicated to frequently asked questions, the startup specifies that the songs generated with the free version of the service remain the property of the company, not of the user who created them, and can only be used for non-commercial purposes. However, the rule changes for users with a paid subscription, who are considered the owners of the generated songs. Regardless of what Suno says, however, there are still no clear rules for audio created by AI. However, last year the US Copyright Office clarified that this is not possible considered protected by copyright visual works produced using artificial intelligence models: there is not yet a specific indication for the production of audio and multimedia, but it is reasonable to assume that the legal ratio can also be extended to musical works.
The other models: Udio, Sonauto, Lyria
Suno is the most promising among the models capable of generating songs, but he’s not the only one. A direct competitor with almost identical characteristics is called Share: It was developed by ex-Google DeepMind engineers and it has already raised millions from some of Silicon Valley’s biggest investors, including Andreessen Horowitz (a16z). And a16z recently suggested in a letter sent to the US Copyright Office that the use of copyrighted materials for training generative artificial intelligence models should be permitted and understood as “fair use”. Fair use, literally “fair use” is the lawful use of protected materials without prior authorization for non-commercial, study and teaching purposesor informative demonstration.
It’s hard to understand why the training of proprietary systems that promise to greatly enrich those who own them should fall under the “fair use” clause. The position, however, does not surprise anyone: Andreessen Horowitz has long been known for its positions techno-optimistic extremism and ultra-liberal of its billionaire founders.
The event
Artificial intelligence, virtual reality, sustainability: what the fashion of the future will be like
by our correspondent Bruno Ruffilli
The reactions of the music industry
The major record labels are of a completely different opinion, it goes without saying. Recently the Universal Music Group (UMG) is suing Anthropic, the OpenAI competitor developing the LLM Claude model, claims that the service generates lyrics from copyrighted songs. Many other legal disputes based on similar allegations, such as the one between the New York Times and OpenAI, are also ongoing in other creative industries, from writing to the visual arts. Their outcome is fundamental, because it could determine American jurisprudence on cases of violation of copyright by AI for years to come.
The problem, however, is not just legal. As already visual artists, illustrators and photographers, even musicians do not share investors’ enthusiasm for the magnificent and progressive fate of generative artificial intelligence models. By early April, more than two hundred artists, including Pearl Jam, Nicki Minaj, Billie Eilish, Stevie Wonder and the Sinatra heirs, had signed up an open letter spread by ARA (Artist Rights Alliance) which harshly criticizes the use of AI for music generation. In the appeal the musicians speak of “existential risk” for the entire sector and ask tech companies and digital distribution services to do not use AI to “violate and devalue the rights of human artists”.
“Don’t get us wrong: we believe that, when used responsibly, AI has a enormous potential to advance human creativity to enable the development and growth of new and exciting experiences for music fans everywhere,” the letter reads. “Unfortunately, some platforms and developers use it to sabotage creativity and undermine the role of artists, authors, musicians and rights holders. If used irresponsibly, AI poses an enormous threat to the ability to protect our privacy, our identity, our music and our livelihoods. Some of the largest and most powerful companies are using, without permission, our work for train AI models. These efforts directly aim to replace the work of human artists with massive amounts of AI-created “sounds” and “images,” which substantially dilute the royalties paid to artists. For many working musicians, artists and songwriters just trying to make ends meet, this has catastrophic consequences.”
#compose #entire #song #spending
– 2024-04-17 00:32:50
