The Center for Data Innovation spoke with Pierre Barreau, chief executive officer and co-founder of AIVA (Artificial Intelligence Virtual Artist), a company based in Luxembourg that uses AI to compose soundtrack music for various kinds of entertainment content. Barreau discussed how humans and machines interact in the creative process and the promising ways in which of AI-generated music may evolve.
Eline Chivot: How did you come up with the idea to combine music and AI?
Pierre Barreau: I once watched this science fiction film called “Her” which features an AI called Samantha. Samantha is a super intelligent virtual assistant that cannot take physical form, and because of that, she decides to write a piece of music that will capture a moment of her life, just like a photograph would. (Editor’s note: You can watch the scene online.)
As someone who was raised in a family of artists and studied computer science, I thought that this idea of musical photographs was fascinating and that’s how I decided, along with my two co-founders Denis Shtefan and Vincent Barreau, to create AIVA.
Chivot: Which technology do you use to teach AIVA music composition, and what does AIVA learn? How do you evaluate and validate AIVA’s original compositions to know that it “sounds right?”
Barreau: We came up with an algorithm that can compose a lot of music very fast and provide demos very fast to clients, which speeds up the creative process and addresses this typical “writer’s block” that composers all find dreadful. We can use our AI to compose emotional music for shows, movies, games, and any kind of entertaining content, but composers can also use the technology for their own creative purposes.
To make this possible, we use deep learning technologies to teach AIVA the art of music composition. AIVA learns music by directly reading 30,000 scores of history’s greatest composers such as Mozart, Bach, Beethoven, Wagner, Stravinsky, etc. From a couple of measures found in existing pieces of music, the AI will try to come up with a continuation and ending for those pieces, and it will compare them with the original scores. If AIVA’s prediction is close and faithful to the original scores, then the AI is getting better at understanding how music is shaped and written.
Although we have developed some automated processes to evaluate how good the music generated by AIVA is, we ultimately evaluate the music ourselves because we are all musicians in the team and some of us are professional composers, so it helps in assessing the quality of AIVA’s creations.
To make sure that AIVA’s creations are original, we use graphics processing unit (GPU) computing to create a “plagiarism checker” which verifies the extent to which AIVA’s original compositions copy the database it learns from.
Chivot: What creative process do you use with AIVA?
Barreau: To know what AIVA should compose, our clients tell us what they need, for instance an emotional or an uplifting song, or if they would like something in the style of a specific composer. Very often, clients will give us a reference track that we will use to bias the training process of our AI. For example, if you wanted a piece of music similar to the Moonlight Sonata by Beethoven, AIVA can create a completely original piece that will have the same emotional characteristics as that of Beethoven.
Chivot: What can AIVA do that human composers and musicians cannot? To what extent does your technology still require human intervention?
Barreau: AIVA can definitely do things that human composers will not do, because it doesn’t have physical constraints. For example, writing a piano piece for 15 fingers instead of 10 is not something that will stop AIVA.
That being said, human intervention is still required to curate and select the best pieces that are generated by our AI, and we actively encourage composers to arrange and modify the pieces that they create with AIVA. We also very much value recording the scores that AIVA creates with human musicians.
Chivot: What are the primary applications you see for AIVA today, and how do you think that may change in the future?
Barreau: For now, we think of AIVA as a great tool for composers to use as a way to get inspiration and to supercharge their creative process. Composers are able to create original scores in seconds with AIVA in a variety of styles, using specific musical parameters.
For the future, we think personalized music will be the single biggest change in the way in which we consume and create music. There are many applications possible for this new type of music. A few examples include: interactive and infinite music for video games; a personalized life soundtrack for each and every individual, based on their story and their personality; using music as a potential cure—or at least a pain reliever—for patients suffering from neurodegenerative diseases; music therapy in general; personalized soundtracks for books that you are reading; and personalized soundtracks in advertisements.