The artificial intelligence research organization, OpenAI, made headlines this week announcing the release of Jukebox, a neural network capable of creating original music and even “rudimentary singing” (or at least something akin to singing). This specific program generates original tunes in a vast range of genres and even emulates the styles of popular artists.
Training Jukebox involved creating a dataset of more than one million songs including LyricWiki lyrical information and metadata, according to OpenAI. This metadata accounts for genre, artist, the year the song was released, as well as moods and keywords associated with playlists featuring these songs.
It’s typical of budding young artists to dabble in a spectrum of styles early in their careers before dedicating themselves to a specific genre and Jukebox is no different. The debut album includes a full gauntlet of genres including pop, country, rock, blues rock, heavy metal, and classic pop. In total, the release features a total of six curated debut tracks, although there’s also a far more extensive uncurated list of Jukebox samples available.
Imitation, as they say, is the highest form of flattery, and if so kudos are in order for the list of artists stylized in Jukebox’s freshman cut. These Billboard regulars include Alan Jackson, Elvis Presley, Katy Perry, Joe Bonamassa, Rage, and even the self-proclaimed chairman of the board himself, Frank Sinatra.
While we were expecting something along the lines of a series of neurons misfiring over a theremin, overall, the songs are fairly impressive. At a low volume, these jams could pass in most environments without raising any eyebrows, however, once you take a more discerning listen or even a slight gander at the lyrics the wheels start to fall off a bit.
To assist, the lyrics in the released songs “have been co-written by a language model and OpenAI researchers.” The lyrics for the most part pass muster aside from maybe a line or two in the Sinatra nod. This song, in particular, opens with: “It’s Christmas time, and you know what that means, Ohhh, it’s hot tub time!”
The overall quality and clarity of the “rudimentary singing” varies wildly from track to track. As noted in an OpenAI release, “singing voices generated by those models, while often sung in a compelling melody, are mostly composed of babbling, rarely producing recognizable English words.”
The Sinatra track sounds more or less like ol’ Blue Eyes. The country ode to Alan Jackson
passes and in all honesty could potentially even inconspicuously slide right in the middle of a few classic saloon hits. Unfortunately, the same cannot be said for the Katy Perry cut. This track in particular fully embodies the aforementioned incomprehensible babbling, almost to a painful degree.
Regardless, the tracks are certainly worth a listen if for nothing other than applauding the impressive milestone in the field of neural networks. Perhaps Jukebox will find it’s genre niche with its sophomore effort, or maybe the AI will continue to dabble in the full spectrum of our mysterious humanity and all of its melodic noises. We will simply have to wait and see…and listen.