Skip to main content

A Guy Invented a Bot That Writes Songs Like Luis Alberto Spinetta

It studied every song the iconic Argentine musician ever composed.

By | [email protected] | April 24, 2019 11:33am


You don’t have to be a Black Mirror fan to acknowledge that technology is rapidly blurring the boundaries of what we thought to be impossible just a few years back. And the arts seem to be in the middle of all the changes taking place throughout society. Filmmakers are casting robots, dead musicians are touring as holograms, and now it seems that Artificial Intelligence may be tackling one of the most sacred pillars of Argentine rock royalty.

An informatics engineer from the University of Buenos Aires (UBA) has actually developed technology in which a bot generates song lyrics that imitate the patterns used by Luis Alberto Spinetta. Freaked out yet? Well, let me explain a little more, to the best of my abilities. But first, let’s listen to El Flaco while we read, shall we?

So what exactly is this all about? Well, this bot is the brainchild of Alex Ingberg, an Argentine that currently resides in Tel Aviv and who published his findings in a complicated, dummy un-friendly post on Medium, which I will now try my absolute best to try and translate to you (wish me luck). It begins, simply enough, with the bot being presented with all of Spinetta’s lyrics since his first album Almendra (1969), to his posthumous work Los Amigo released in 2015. It took into account all songs in which he had at least a partial credit in composing. So far so good, all is understood. But this is when it all gets tricky.

Ingberg created two different algorithms. The first, created from a process called Markov chain, which didn’t yield very original results, as the mechanism works based on probabilities. In Ingberg’s own words, the result was “copy-paste of [Spinetta’s] work, a collage among different lyrics to create a new song,” and it even dished out several phrases that were identical to those written by El Flaco himself. But the second algorithm is where the magic really happened. In it, Ingber worked with something called Recurrent Neural Network (RNN) which recognizes patterns in data streams and also adds the temporal dimension. Which means it kinda, sorta remembers… I know what you’re thinking:


After this simple introduction, comes a very complicated explanation that uses terms like LSTM, Python 3, and something called Greg Surma’s text predictor (again, if you wanna go deep into this go to Mr. Ingberg’s original post). Suffice it to say that after 50,000 iterations he was able to obtain some real, palpable results. In his words again: “The interesting thing about this is not just seeing the final result, but the learning process that the algorithm was having in every iteration. From a concoction of characters, it goes on to create formed verses in only a few cycles.”

The results are, nevertheless very difficult to evaluate by an untrained eye. They’re full of grammatical and even spelling errors, and in some instances they just don’t make any sense at all. But for Ingberg, the findings couldn’t be more of a success. “Although the lyrics generated by artificial intelligence have these little flaws, we can definitely see that the model correctly learned to copy the style of the provided dataset,” he says. In other words, the results are quite extraordinary, considering that the network never received any data from a dictionary or a grammar book. Just El Flaco’s words. No news yet about any Spinetta hologram being developed for next year’s Lollapalooza, though.