Here’s a doodle that doesn’t work at all. It celebrates Johann Sebastian Bach’s birthday (21 March) with what at first sight looks like a surprisingly sophisticated device. A mouse click unpacks a musical box that allows you to write a melody of two bars. The “machine” will then harmonize it in “Bach’s signature music style”.
Let’s ignore those silly, plasticky playmobil puppets, the cheap synth sound, the dreadful on hold music (no trace of Bach there!), and stick to what it claims to do: synthesizing Bach’s harmony by the application of Artificial Intelligence.
And let’s accept that it has been trained upon chorale melodies only. You can’t expect it to do what Bach would have done, if forced in purgatory to harmonize Eine kleine Nachtmusik or Rock around the Clock.
Even with melodies that should fall within its scope, the result is gibberish. Here are two harmonizations of one chorale-like melody. (Click images for sound.)
In defiance of its training corpus of chorales, the machine produces an abundance of precisely those things God and Bach forbid: parallel octaves, unsingable jumps, false doublings, erratic dissonances, and so on. Mistakes that not even a moderately musical human beginner would be tempted to make.
So, how was it made to learn (or not learn) the rules of Bach’s “signature style”?
Most of my harmonies follow a standard structure, making it fairly easy for a machine learning model to identify patterns in it. Take that, Beethoven!
So Playmobil Bach himself tells us.
The patterns of Bach’s harmony are, in fact, famously complex. But even the most basic rules are not mastered by the machine (such as avoiding parallels of octaves and fifths). The machine can’t have found many of those in its training corpus. Nor has it developed a clear sense of what it means to harmonize within a specific key (C major, G major, etc.).
Machine learning models need time to learn how to find the right patterns. This machine trained for a few hours, despite the relatively small dataset of 306 Bach compositions.
Which implies that a few hours is a long time. It looks like the team that has produced Playmobil Bach has underestimated the complexity of the task.
And lacked basic musical expertise – it is hard to imagine that professional musicians or music theorists have been involved in the project. No musician would propose to generate a harmonization from the top down, first alto, then tenor, bass, as the machine shows us.
Research into artificial musical intelligence has been under way for several decades. Algorithmic composition has a history of centuries. It has produced results that are mostly either trivial, or feeble copies. But even so, Playmobil Bach is far from state of the art. It’s a guessing machine, with no traces of intelligence.
And that’s a disappointment. It could have revealed something of the inner workings of music to mass audiences. Imagine what could be done if Google would invest a little of its wealth and power in music education. Seriously.
Take that, Google!