Am 17. Juni findet das Creative AI Symposium statt, bei dem wir gemeinsam die Möglichkeiten von kreativer KI erkunden. Einer der Referent:innen ist Dr. Tristan Behrens, der mit Künstlicher Intelligenz Musik generiert. Im ARIC-Interview spricht er über KI im kreativen Prozess, die passende Datenbasis für KI-Komposition und wagt einen Blick in die Zukunft der KI-Musik.
ARIC: You describe yourself as an AI guru. What does that mean?
Dr. Tristan Behrens: A colleague called me a deep learning yogi because I am a yoga teacher, but the domain deeplearningyogi was too long for me. In Eastern philosophy, a guru is someone who brings light into the darkness. And I try to shed some light on AI through my teaching.
Is AI a dark something in the sense of huge and unmanageable?
It takes time to get into it. If you ask “What is the definition of AI?” in a room with a hundred people, you’ll get a hundred different answers. In my day-to-day teaching work, I try to make sure – just like in yoga practice – that people at least do a handstand. Because we’re at a threshold right now: We have done a lot of research and are now taking it into the industry.
We invited you to the Creative AI Symposium because you are involved with AI-generated music. Where does your interest come from?
When I was studying computer science, I was bored during the semester break and then I started producing music. Music accompanied me for years and I learned programming anyway. In 2017 or 2018, I got the inspiration that I could do both at the same time: Programming and then add music to it.
And I have no regrets: it’s inspiring to the max.
What do you find inspiring about generating music using AI?
There are different phases in the creative process. There is the phase where you have to find an idea that leads to a high quality. Today, AIs can inspire so well, almost at the touch of a button, that you usually have too many ideas and are then faced with the question: Which one is the one you really use?
Does the AI generation work better or worse for certain music styles?
All styles work equally well if you have a sufficiently large amount of data to train with.
“If I only train an AI on the works of J.S. Bach, no Rammstein will come out.”
In other words, there are only indirect problems due to the database?
If I train an AI only on the works of J.S. Bach, no Rammstein comes out. Recently, however, I trained an AI on everything, with around 400,000 songs. The result is an AI that can be used universally. It can produce chamber music, but the same AI can also produce rock’n’roll.
What kind of database do you need?
I work with symbolic music that is already encoded as notes. So it’s not the sound files. The equivalent of “Hallo Welt”, which I offer people on my site, are Bach’s chorales. That’s a maximum of 400 songs. You can do something with that. The result is wonderful music.
If you have 1000 or 10,000 songs, you already have a good basis.
Our workshop lasts 1.5 hours? How much can you actually accomplish in such a short time?
I planned to introduce people to the principle in the first 1.5 hours: Predicting the next grade, what the data should look like so that you can feed it to an AI, what the models look like and how to train the model. I have prepared a data set and the architecture. We will complete this together and train an AI.
How has it changed in recent years?
Computers have become faster. We can run ever larger neural networks on our relatively small hardware.
The software is now also so easy to use that there are only low barriers to entry. One example is the Hugging Face platform, where you can train a language model with very little source code. Just a few years ago, eighty percent of the time was spent on the data and twenty percent of the time on the neural networks. That has shifted: Now you spend ninety percent of the time on the data. That’s brilliant! The time it takes to get results has been reduced.
What could you imagine? Where will we be in ten years in terms of AI and music?
At the moment, it still takes a certain amount of effort to operate AI music, but it is slowly coming onto the market and the big music companies are following suit. In ten years at the latest, people will be using AI-supported tools to compose beautiful harmonies and generate songs. What’s more is the timbre, the timbre through AI instruments. I would be able to create my entire work based on text: Sound, artwork, description, music videos – and probably on our little computers.
What are the commercial applications?
Isotope is one example: the company uses deep neural networks. There are also providers for mixing and mastering that have an AI behind them. It listens to the music, compares it with a reference mix and then creates its own high-quality mix. After a minute, you have a nice mix. Work that used to take a lot of time is made easier by these providers.
What reactions do you actually get to your work?
When I play my work, i.e. my music, I see people bobbing their heads. So I almost only get positive reactions.
Falls unser Interview mit Tristan euer Interesse für KI und Kreativität geweckt hat, laden wir euch herzlich zum Creative AI Symposium 2022 am 17. Juni ein. Der Workshop mit Tristan Behrens ist nur einer von vielen spannenden Programmpunkten. Ihr könnt einen der Praxisworkshops belegen und z.B. lernen, wie man mit KI schreibt und akademischen Vorträgen lauschen - zum Beispiel zum Thema AI and Quantum Computing. Hier klicken für mehr Infos. Hier geht's direkt zur kostenlosen Anmeldung.
