Je secoue le monde (I Rock the world)

VOTE FOR US
A.I Song Contest

Listen to the song
on YouTube
on Spotify
on Deezer

Lyrics written by an AI (GPT-2 based) :
Je secoue le monde
Oh, ne m'appelle pas
Ne m'appelle pas
Oh Oh
Ne m'appelle pas
Ne m'appelle pas
Oh
ne m'appelle pas
ne m'appelle pas
Oh Oh Oh
ne m'appelle pas
drôle de lapin

Je suis Robert de Baratin
Bonjour

Je secoue le monde
Let's sing along

Ne m'appelle pas
Oh OhOh
Ne m'appelle pas
Drôle de lapin

I'll blow your money, money
I'll get you to my bad ass spinning' for you
Oh ne m'appelle pas drôle de lapin

Oh, I'll make it easy, honey
I'll take your money, yummy, yummy
I've got my bitches standin' up next to me

Oh, ne m'appelle pas drôle de lapin

So come and check it
Alors venez le vérifier
Mon mauvais cul tourne pour toi
My bad ass spinnin' for you

Mon mauvais cul tourne pour toi
Oh ne m'appelle pas
drôle de lapin

Même lorsque le tonnerre et la tempête commencent
Je serai fort comme un arbre dans le vent
Rien ne va bouger cette montagne

Je tombe du ciel
Et je suis tout seul

Je tombe du ciel
Et je suis tout seul

Tout seul
Tout seul

Oh Ne m'appelle pas
Drôle de lapin


Interview by Karen AI Contest Song

Anything you want the listener to listen for?
Yes of course ! We hope that people will like the sexy and languorous voice of Robert de Barretin!

Why did you decide to take part in the AI Song competition?
We are a committed artistic collective. We create offbeat artistic pieces using data and A.I as raw materials. We are approaching the subjects of data and artificial intelligence from a new angle. The AI ​​Song Contest was a great opportunity for us to explore open source bricks and AI creation tools. We used them to create a totally DataDada song based on a dataset of lyrics and music bringing together all the songs from all the competitions of past Eurovisions.
Can you describe your approach and creative process for the song Je secoue le monde? Our first step, the most important for people mad of data, was to consolidate and enrich the data set provided for the competition augmented by the ones of Spijkervet available on github.
We took the time to discover and identify the specifics of this “Eurovision” dataset. Then, we enriched it and defined guidelines for the creative work ... Because with AI, everything is in the dataset.
Following this dataset analysis phase, we decided to focus only on the Eurovision winners, ranked first, second or third. Because DataDada loves winners, and only winners! Do you know that in 1969 the five finalists received the same score and all finished first. Unbelievable, right?
So we looked for all the original lyrics from all of the Eurovision winners; some, in original languages, have been massively translated thanks to our friend Google Translate. On the music side, we selected the winners' midi files by identifying them from the original provided dataset of midi files. No other data sources have been added. When we love data, we care about having a 100% Winner dataset!

We then isolated the tracks of each midi song to make a dataset consisting of tracks 1, tracks 2, tracks 3 and tracks 4 of all the songs.

Lyrics Dataset is available here
Midi Dataset is available here
So, as you can see, we have worked a lot on the dataset, favoring data from Eurovision winners.


What was the inspiration for the song?
We didn't have a particular inspiration, but rather a data-oriented angle of attack. The song created can be seen as an interpretation of a dataset with a hint of creation. The final song is both goofy and surprising. It was essential for us to make our DataDada spirit emerge from a rather classic, or even fairly formatted, songs dataset, as Eurovision is.
Always very fond of playing with words, generating texts from GPT-2-simple was a real treat because it was full of surprises.


What are the main assets and tools used to create the song “I shake the world”?
After consolidating the dataset, we handled open source AI bricks like GPT-2-simple, Char RNN or even Magenta studio for Ableton. This learning stage was very fruitful.
For the voice, we used our own voice assistant Robert de Barretin. Robert, despite his male first name, has a female voice. You can also play with him (in English but with a French accent).
Finally, we mixed in the background the voice of Albertine Meunier to sing with Robert. We used a very powerful Alien computer with a 1080 GPU to run machine learning.
We installed GPT2-simple and trained the model with our lyrics dataset of the Eurovision winning trios. Once the model was trained, we were able to generate wonderful lyrics!
Here is a small overview of the lyrics generated by the GPT2-simple program fed by our Eurovision dataset.
We chose our favorite among all, namely the lyrics of the text I_Rock_the_World_ # 8. Despite numerous attempts, we were unable to install Music VAE from Magenta. We encountered problems during the installation of Music VAE not compatible at that time with the available version of Tensorflow.
Also, we decided to change and opted for Char RNN and Magenta Studio for Ableton to generate our MIDI tracks for music.
Char-Rnn uses the transformation of midi to Anglo Saxon writing of music, thus creating sequences of words made up of combined letters. ABCDEFG.
Then, we generated midi tracks from this dataset and the wonderful Eurovision midi generation tool developed by Bastien Didier.

Magenta studio for Ableton was our solution to create the rhythm of our song. We spontaneously created a random drum rack and then ran Magenta in Ableton to produce musical loops in .wav format.
Here are some results.

We also installed Deezer's Spleeter to separate the soundtrack from the voice in mp3 files. We ran Spleeter for the song "la vie en rose", which recently fell into the public domain because it was created in 1950. But in the end we did not use this material for our song. However, you can listen to the separated soundtracks.


How did you decide which parts of the songs to create by hand and which parts to “machine”?
From our point of view, everything is in the data and in the dataset ... so we have worked a lot to have a consistent and quality dataset.
Everything is also in the curation of the dataset and in the choices made for the creation.
These choices were decisive:
1) Take only the data from the songs of the Eurovision winners
Handmade

2) Start our song with the first sentence “I rock the world”. The choice of this expression "I rock the world" is based on the frequency of the words of all the lyrics in English.
Machine made with word count program

3) No generation produced by our AI models has been transformed, in this sense the production is gross of I.A production.
Machine made

4) The choice among all the generations of texts and midis and wavs was made in a completely intuitive way, none of us having musical knowledge.
Handmade

5) The arrangement of the midi and wav tracks was done in Ableton Live Handmade


What was the main challenge faced by your team?
Before starting this project we had no experience with the AI software AI, and we felt like great beginners of AI lyrics and music. But we trusted AI and our ability to choose from the mass of machine productions.
Our main motivation was to create a DataDada song. And this, without the intervention of any musician.
For the arrangement of the song, we asked Lily Gray, singer and informed user of Ableton for her own compositions, to help us.


Why did you start on “I rock the world” / “je secoue le monde”?
GPT-2 text generation programs need a text beginning. To define it, we chose from the most used words in the dataset of all the Eurovision songs we had at the start of the project. t As you can easily predict, the word "Love" comes in first position counting all Eurovision songs. But we chose "I rock the world" because on the one hand in French Je secoue le monde "is quite surprising as an expression and on the other hand this title is quite adapted to our era totally shaken and "confused "!


What was your criteria for creating a hit song?
Our main criterion was to create a song with the DataDada spirit!


Photo & LP design Frédéric Mit