While us human beings were trying to get into summer shape, our algorithm also got swimwear ready. What happens when you send a self-learning algorithm to a non-human summer training camp, you ask? Let’s head over to our Research department to find the answer. Warning: Chordifying a new song may never feel the same again after reading this article.
Did you know that 99.9998% of all the songs you find here at Chordify are non-human annotated songs? Yup! It’s top notch machine learning technology that runs the show here, almost no mortals needed anymore. It just needs some proper training every once in a while. Before you find out what this computer-science-based-bootcamp is all about, we need to explain what our algorithm is exactly. Since most of us didn’t study Artificial Intelligence, this is a highly simplified explanation.
What happens when you hit the ‘Chordify now’ button?
What goes on under the hood when you search for a song on Chordify? If the song is already chordified, the song’s chords are pulled from a shelf out of the database that stores the chord progressions for millions of songs.
If the chords are not in the database yet, you’ll find the ‘Chordify now’ button. Hitting that button kickstarts the algorithm engine that immediately starts listening to the song to calculate the chords based on the audio signal. This means that behind the scenes, there are some deep neural networks constantly listening to millions of songs and spitting out an endless stream of chords that are shown on the right point in time. To do this, you need a beat. Or to be more specific: a beat tracker. The first time we presented this beat tracker was at ISMIR 2019, where we showed it to scientists from all over the world.
Down beat tracker
So, the Chordify algorithm is based on deep neural networks; a chord network and a beat network. Together, they make the chords appear at the right time in the song. The beat tracker makes the chords hop along with the music on time, just like you nodding along with the music, but then without the need of a head. What’s special about this beat tracker, is that it not only nods along, but it also detects the beginning of each bar, marking it 1 (2, 3, 4). The 1 is called the down beat. The training we did this summer was all about improving the down beat tracker.
4238 songs in 4 days
As you know now, the Chordify algorithm works its way through thousands over thousands of songs every day, running its laps on a steady basis. To enhance the quality of the algorithm (in other words, the accuracy of the chords and beats) – that’s where the training comes in! Together with one dedicated server that ran non-stop for four days, the deep neural network learned from 4238 songs during this bootcamp.
Man versus machine
The learning process involves a comparison of the network outputs to what actually should come out, based on human-made annotations. In other words; to what extent does artificial intelligence match human intelligence? After that comparison is made, this information goes back through the network in the opposite direction and alters all the nodes in the network such that the network gets closer to output the right thing in the first place. This summer, it learned to track the down beat more accurately.
Better down beats for all
Now that you understand the basis on deep neural networks and where the quality improvement of the algorithm comes from, hitting the ‘Chordify now’ button might feel somewhat more magical to you. You might not notice the improvements every time you use Chordify, but it’s these technological details that do the trick in the long run: improving our algorithm and the chords it provides, so you can play every song you want. Happy jamming!