Is it even possible for a machine to produce original music? What’s the closest that we’ve gotten today? And what’s creativity anyway? In this post, IT experts from Sirinsoftware.com will attempt to answer these questions.
To better understand AI-generated music, we should first look at how music is created by humans. So, let’s jump straight in!
How Do Humans Create Music?
When we think about music creation, it usually comes in two main forms: playing live instruments and tracking them in software or using digital instruments and arranging them also in software. At the genesis of either method, a song usually begins with a ‘seed.’ And that seed can be anything from playing a guitar riff, a bass line, some synth chords, or even a vocal sample. When a musician listens to a seed, they begin to hear what next comes to their head, figure out how to get what’s in their head into the computer, and then layer it on top. Composers also repeat the process until some tune or melody begins to take shape. This is just how some musicians can imagine their music creation process. It may be different for others, though.
How Do Machines Create Music?
Needless to say, artificial intelligence doesn’t rely on the same patterns and music creation methods as its human counterpart. Neither does it need inspiration or creativity to generate tunes. Machines listen to many examples and determine some patterns that they will be able to utilize to create something new. AI doesn’t need to consistently receive put from a programmer. It just builds on the previously mastered patterns and learns from its so-called experience. For example, a human user can input two or more types of melodies and then use machine learning to combine them in a new way.
AI-Generated Music Today
And now, let’s take a look at the state of the art in AI creating music.
So, currently, machines can create music in two main ways: either manipulating mini-data or raw audio synthesis. Let’s check out the best examples of both.
Google’s Magenta is the music writing AI that successfully learns from previous melodies, drum patterns, and other sounds to create new ones. In 2021, a Toronto team created the Lost Tapes of the 27 Club, a project featuring songs written and mostly performed by machines in the styles of other musicians who died at age 27. Among some of the famous performers were Amy Winehouse, Jim Morrison, and Jimi Hendrix. Sony has previously used software to make a new song from the Beatles. So, here is how it works. Each track is the result of AI programs analyzing up to 30 songs by each artist and granularly studying the song section by section, that is, drums, guitars, vocals, and more. For example, if you need a new guitar riff, you just need to input a bunch of guitarists from the original artist, and then AI will come up with its own masterpiece.
The limitation in all of this though is that it can’t do the entire song’s structure, or it gets confused. So, digging deeper rather than the artist’s song being fed in and processed as raw audio, they are first converted into MIDI files. MIDI files are basically bits of code with information that tells the computer exactly how to play a digital instrument, that is, the volume, the length of a note, the particular beat or rhythm, etc. After examining each artist’s note choices, rhythm process, and preferences for harmony in MIDI format, AI creates new melodies in the form of MIDI files. At this stage, it’s a jumble of notes. And it’s up to human artists to sift through and pick out the best moments. By using artists’ previous lyrics, AI also tries to come up with new lyrics in a specific artist’s style. But once again, these have to be sifted through by humans to make sense.
Once the raw MIDI files and lyrics are in place, they can be performed by cover artists. For example, Erick Hogan performed as Kurt Cobain. The Google team also created Ableton plugins, which allow AI to listen to a bass line and automatically generate a drum pattern to go with it.
Open AI has taken a different approach. Instead of doing MIDI files, it uses raw audio to train a model, which, in its turn, spits out raw audio. And most musicians think this method of generating music is more interesting. The models are trained on a raw data set of 1.2 million songs and used metadata and lyrics from Lyric Wiki. The program works in two main ways: specify a genre and this will make something from scratch or feed it a section of a song and let it continue writing that song.
There are lots of other fun and interesting approaches ubiquitous AI uses to generate music, which has all the chances to sound as good as those created by humans. Just keep on exploring to learn more.
Hi, I’m Erick Ycaza — a music blogger with a BA in Advertising & Graphic Design. I created this blog to keep you updated with daily music news. Surprisingly, I’ve been writing about music since 2007. If you’re an artist and would like to be featured, feel free to reach out: info@electrowow.net
