As Christmas is fast approaching, we are starting to hear the holiday classics everywhere. From All I Want for Christmas at the Supermarket to Rockin’ Around the Christmas Tree on the radio – Christmas songs are virtually unavoidable.
But I’d like you to think of the artists behind these songs – most of Spotify’s Christmas Hits playlist[1] is comprised of songs recorded or written before the first manned mission to the Moon.
Naturally, many of the authors and performers listed in the credit sections of these songs are long gone – Bing Crosby died in 1977, Nat King Cole passed away in 1965 and Frank Sinatra departed in 1998.
It’s a shame that we won’t be able to hear any new songs from them.
But what if it doesn’t have to be that way?
That’s where OpenAI’s Jukebox comes into play.
Debuted in April 2020, the technology analysed over a million songs[2], along with their lyrics and metadata (release date, genre, mood) and is now capable of generating full tracks in the style of any well-known artist. The company shared a range of demos, designed to resemble artists such as Alan Jackson, Katy Perry, or Elvis Presley. Most notably though, the song that stands out is “Hot Tub Christmas”, in the style of Frank Sinatra. While the “recording” quality might not be perfect, the timbre of the “singer’s” voice is eerily similar to that of the legendary American singer.
Though the lyrics have been co-written by a language model and OpenAI researchers, the chord progressions and instrumental cohesion are very well replicated in the computer-generated mp3s. The team behind Jukebox is aware of the software’s faults, as “[…] the generated songs show local musical coherence, follow traditional chord patterns and can even feature impressive solos, we do not hear familiar larger musical structures such as choruses that repeat.”2
Jukebox doesn’t analyze the actual notes in the songs, but only relations between pitch over time. An upside of such an approach is the possibility of highly realistic human voice creation. For their future endeavors, OpenAI plans to integrate a note-to-MIDI technology which would detect the rhythms and exact notes, which would allow for a deeper, more natural, and precise song creation – perhaps with the use of software instruments or synthesizers for higher file and sound quality.
Jukebox, at this point, is treated by the music industry as a mere curiosity, with no real applications – even despite a new feature of creating an acapella file from user-generated lyrics being introduced. This dynamic might change in a relatively short time if Jukebox becomes able to create classically written songs, providing the notes, rhythms MIDI files behind them. With such possibilities, songwriters and producers could streamline their music creation processes and massively increase their output.
The current market situation is visualized by the fact that most of the investments poured into creative AI come from Venture Capital and Tech Corporations – not from the Music Industry.[3]
At this point, it does not seem like any songwriter or producer jobs are endangered. High quality audio files have incredibly many timesteps which encode data – a standard 4-minute-long song in a .wav 44.1 kHz file will contain over 10 million timesteps.[4] Currently, a song needs to be almost fully produced and designed by a professional before being rendered into such a complicated audio file.
The process of AI art generation is slowly being integrated into commercial culture, with the generator Midjourney winning the Colorado State Fair Fine Arts Competition.[5] Jukebox and similar technologies are often criticized for taking away the humanity out of art, while some perceive it as an opportunity to augment their creations through technology.[6]
To me, it seems inevitable that Artificial Intelligence will be widely used in the music industry – major labels will push for anything that can give them a competitive edge in business.
We must also take into consideration the legal implications of Jukebox.[7] Our laws don’t include AI “artists” and thus, there might be copyright implications. Who is the de facto author of such a song? The AI developer, or the person who entered prompts into the technology to create a specific tune? How do we split royalties for such songs? Furthermore, is it acceptable ethically to expand dead artists’ catalogues?
In conclusion: AI is slowly entering into creative arts, especially the music industry, thus expanding songwriters’ and producers’ output and possibilities. It appears that in this case, the risk of actual people being replaced by technology is lower than in easily automated and routine operations.
This time, I’ll ask the classic question: do you think that AI art is proper art? Should it be publicly disclosed that a song or a painting was generated through Artificial Intelligence?
Let me know what you think in the comments!
Until next time,
Jan
[1] https://open.spotify.com/playlist/37i9dQZF1DX0Yxoavh5qJV?si=d4fa601b2c3f4418
[2] https://www.cnet.com/science/these-ai-generated-katy-perry-and-elvis-songs-sound-hauntingly-real/
[3] https://blog.songtrust.com/current-state-of-ai-what-songwriters-need-to-know
[4] https://openai.com/blog/jukebox/
[5] https://edition.cnn.com/2022/09/03/tech/ai-art-fair-winner-controversy/index.html
[6] https://fortune.com/2018/10/25/artificial-intelligence-music/
[7] https://themix.musixmatch.com/post/ai-in-songwriting-4-practical-applications