Part 2

Tell me about the compositional process and how the music gradually took shape, please.

It was quite a long and empirical process over three years. I tend to compose through research rather than starting with a completely defined idea.

One element that really helped conceptualise the direction of the album was vocal synthesis, particularly the rapid development of voice cloning around 2022. I had already been experimenting a lot with different kinds of vocal synthesis and processing, but suddenly there were new possibilities that really excited me.

At the same time, I continued experimenting with MIDI scores. I might start with just a few bars from one score, combine them with another, isolate them, slow them down, transpose them, loop them and try completely different sounds. At some point, a combination produces a particular emotion.

Then I accumulate a lot of material, almost like rushes: synthesis, processing, drones, melodic fragments, textures, voices, accidents. I don’t really separate composition and sound design. A timbral decision can completely change the perception of a chord or melody and make me rewrite the composition, and the reverse can happen as well.

Gradually, I noticed I was using more sounds that evoked strings, or actual string samples. That led me to develop pieces where real strings could coexist with very electronic sounds and completely digital voices. I started writing fairly complete string arrangements using synthetic sounds and multi-sampled instruments/ virtual strings, like you might do in film music before recording. The real strings then replaced or interacted with those elements, not as a classical layer added on top, but as another material.

I had also wanted for a long time to make a more substantial artistic project with Echo Collective. I’ve worked with them for around ten years as an mixing and recording engineer, producer, so that happened quite naturally.

The final form of the tracks emerged progressively. I think a lot in terms of density, energy and contrast rather than traditional structures, moving from something very fragile to something excessive or almost overwhelming.

How did you go about recording the string sections? Did the musicians hear the rest of the arrangements or did they know anything about the other parts of the composition? Do you think these things matter?

The strings were recorded as a quintet with Echo Collective. The arrangements were already fairly developed, with virtual strings sitting among the synthesizers, voices and the rest of the production. Neil Leiter, Fabien Leseure and I then worked together on the string arrangements to structure the score and define the playing indications for the recording.

The musicians would hear enough of the full track to understand its overall energy, but during the actual takes we could focus mainly on the strings. I think knowing the context definitely affects the initial interpretation. But once the musicians understand that energy, I don’t think they necessarily need to hear the entire arrangement during every take. Sometimes a click is enough, as long as we can make sure from the control room that the performance still carries the right energy.

Then, during production, that material can be edited, displaced, transformed or combined with elements whose final form the musicians couldn’t fully anticipate. I like that meeting between the two.

Alongside strings, the arrangements also integrate AI voices. How did that work in practise? Were you working with very precise or rather deliberately open-ended prompts?

There are actually no prompts involved in the process, and it isn’t AI generated music or sound. I used AI powered voice cloning, which in my practice feels much closer to a new form of vocal synthesis.

The input is real human audio, including experimental recordings of my own voice and sometimes other recordings. The model can then reproduce that performance with another vocal timbre or sometimes in a completely different register.

I started experimenting with these techniques around 2022, when the systems were still quite unstable, and I was very interested in their mistakes. Pushing a model into a register that was too low or too high for it could generate strange artefacts, shifts in timbre or completely improbable articulations.

Once I find something interesting, I go back to fairly traditional production tools: Auto-Tune, Melodyne, pitch shifting, time stretching, distortion, editing or resampling. I might cut a few syllables, reconstruct a phrase that was never actually spoken that way, then feed that new material through another model and obtain another voice with a new set of aberrations.

I’ve always been interested in vocal synthesis , vocoders, extreme pitch correction, formant synthesis, wavetable techniques, and softwares such as Alter Ego. So when voice cloning reached this point, I saw it as another step in that history, and it became quite important in shaping the album, comme un fil conducteur.

Why did you prefer to work with AI in this case?

All the voices actually begin with human recordings, so for me it isn’t really a true opposition between human and AI voices in the production process. What changes is what happens to the voice afterwards.

Once it has passed through voice cloning, been cut up, reconstructed and transformed, it gradually becomes detached from the body that originally produced it. I can chop tiny fragments almost like samples, build endless phrases or verses from them, and then pass that construction through another model to give it another timbre, whith his new artefacts and strange language bugs.

At that point, you still hear something that strongly suggests a human presence, but there is no longer an actual human performance corresponding to what you hear. It wasn’t about finding a shortcut or avoiding singers. But this process creates another kind of ambiguity between the source and the result.

Conceptually, that worked very naturally within the album. On one side, I’m using old scores whose original sound has disappeared and interpreting them with contemporary tools. On the other, I’m using traces of human voices whose original identity and performance gradually dissolve through digital processing.

I found one description of the piece very interesting: “the hesitation between material and simulation present in the music.” What does that mean?

There are constantly very physical elements such as the strings, but alongside them, there are synthesizers and synthetic voices that can sometimes sound almost more organic than the acoustic elements. And the reverse can happen as well, a real string recording can be edited and transformed until it starts to resemble a synthesizer.

I wanted that boundary to remain unstable, so you don’t always immediately know what was played, simulated or transformed. There is also a parallel in the cover artwork by artist Aline Bouvy. It is actually a photograph of a resin cast of my head, but through the image and the way it is lit, it becomes difficult to tell whether what you are looking at is organic or 3D-generated.

It also feels quite natural to me in contemporary music production. The distinction between an acoustic instrument and its simulation is increasingly blurred. You can record a real instrument, sample it, resynthesise and transform it, then combine it with a virtual instrument that is itself trying to reproduce something real.

Something can feel familiar, human or almost ancient, and then one detail suddenly shifts your perception somewhere else.

By removing the human body from the equation, the AI voices are said to have become “poetic material.” Amidst a lot of negative comments about the use of AI in the creative process, tell me what, on a more fundamental level, AI can contribute to the expressive or creative potential of music, please.

I think it really depends on how you use it.

Personally, I’m not particularly interested in asking an AI to make a track for me or directly reproduce another artist’s work. But I can imagine that, for some artists in the future, generation could be used to create raw material, almost like rushes that can then be chopped, sampled, recomposed and diverted into something else.

In some ways, these kinds of uses raise questions that remind me of the early debates around sampling, without saying the situations are identical, particularly around copyright, sources and the reuse of existing material.

In my case, voice cloning interested me more directly because I see it as an important development in the history of vocal synthesis. These tools were initially designed largely for applications such as dialogue replacement in film and post-production. I was diverting them into something else. I think part of artistic practice can be to appropriate new technologies while also questioning our relationship to them.

Interestingly, by the end of the album production, the models had already improved dramatically and were becoming almost too perfect. I was much more interested in the aberrations, mistakes, strange timbres and invented languages produced by the earlier systems.

At that point, the technology simply becomes another instrument for me. It becomes artistically interesting when it shifts our imagination or perception, rather than simply automating something we already know how to do.