Artificial Intelligence

Artificial Commonsence

Practificial Frankinsence

Beneficial Unintelligence

Machine Learning, the buzz word. Since I started on the MCT master, I’ve been looking forward to these two intensive weeks of coding. I imagined building a whole choir of cute singing robots, performing Beethoven’s 9th symphony:

Or to create a music buddy, who could be my new band mate, something like Dr. Squiggles:

Or an android guitarist with 78 fingers. This is what I need in my band obviously. My Guitar Pro compositions from 2007 (with guitar tabs that demanded the guitarist to have at least 3 hands with 7 fingers each), could have it’s renaissance:

Robot guitarist Image source:

So what’s machine learning? It can be many things. It can be something that learns, after you give it a lot of data with labels attached, and after some training, it tries to predict something (supervised learning). Or it can be an algorithm that on it’s own organizes loads of data into lumps, without labels, and it’s called clustering or unsupervised learning. It can be something that replicates the network of neurons in a human brain: Artificial neural network.

So Machine Learning is a lot of things, really. It’s not just the robots that are about to get sentient (and soon will take over the world etc). And if you get really good at machine learning, you can make a youtube channel that live streams deathmetal 24/7 until infinity:

(This is by the way, what we all will be listening to in the future, when the robots take over the world)

If you are not THAT good at machine learning, but had a two weeks intensive MCT Music and Machine Learning course, you could for instance do something like I did: make a system that identifies which vocal technique you’re using! Why, you might say, don’t we have vocal teachers that do that? Well, if this corona hell is going to last, we might really need some online tools for music instrument learning assessment. If your internet connection is bad, and you sound like Dadabot’s Death Metal livestream through Skype, what should your vocal teacher say, I mean. Therefore I present:

The Vocal Technique Classifier:

To make this tool you need:

Before you pour the whole datset into the algorithm, it should be processed with an extensive procedure of feature extraction. Add 1 MFCC and 1 Melspectrogram with Librosa. Now it’s time to add the emulsifying agent, the Linear Discriminant Analysis. Mix it all together until it turns to a delishious spaghetti code and pour it all into the Support Vector Machine (Optional: Artificial Neural Network). Turn on your CPU to 250 degrees celcius and wait for 5-10 minutes. Bon Apetit.


The delightful chaos before applying LDA:

Before LDA

The beautiful tidiness after applying LDA:

After LDA

Unicorn spaghetti in 450 dimentions:

Unicorn Spaghetti

Spectrogram before LDA:

Spectrogram before LDA

Barcode of vocal features:

Spectrogram after LDA