Send a sound. It becomes part of what the system can say.
What this is, plainly. This is a machine-learning electroacoustic composition project, and what comes out of it is an album.
You send a recording. It goes into a corpus — a sound database, cut into slices and analysed. A machine-listening system built at IRCAM plays that corpus in real time, responding to what it hears. I perform with it, and shape those sessions into finished pieces. Those pieces are the album.
So: not a compilation, not a sample pack, not a remix project. Your recording is raw material for a piece of music that neither of us could write alone.
I am making an album called What the C*rp*s. Its material will not come from a sample library. It will come from people who agree to be in it.
This site is the invitation, the instructions and the place to upload. It takes about ten minutes to take part, and you do not need to be a musician.
Deadline 30 November 2026. The album is released on Bandcamp in February 2027, and every single person who contributes a recording is named in the credits — however small the contribution.
What a corpus is
A corpus is a body of recorded sound cut into slices and measured one slice at a time — pitch, brightness, spectral shape. It is not a playlist and not a sample pack. It is the pool of material a listening system is allowed to reach into.
I work with Somax2, developed at IRCAM. It listens continuously to an incoming sound and, in real time, searches the corpus for slices that resemble what it is hearing, then plays them — in an order the corpus never had. It does not play your recording back. It takes its shape apart and answers in whatever else is in the pool.
So the corpus is the instrument. Whatever goes into it is the full extent of what the system can say. Right now I am building it, and that is what I am asking you for.
What happens to your recording
- CutYour file is sliced at its attacks. Each slice becomes one usable fragment — roughly half a second to five.
- AnalyseEvery slice is measured: pitch, pitch class, spectral shape. That measurement is the only way the system knows the slice exists.
- NavigateA driving sound is fed in live. The system matches it against the corpus and plays the closest slices it can find.
- FixOne continuous pass is recorded, then edited and mastered. That take becomes a track.
Two rules fall out of steps 1 and 2, and they are the reason the recording notes say what they say: a slice with no attack never gets cut, and a slice with a chopped tail gets measured as something it is not. Dry, spaced, and left to ring. Everything else is taste.
What I am asking for
Recordings. They do not have to be instruments. If you make sound with objects — metal, wood, glass, paper, stone, water, springs, boxes, junk, whatever is in the room — you are not a second-best contributor here. You are exactly who I am asking. Some of the most useful material in a corpus has never been near a concert hall.
Extended techniques are just as welcome: bowing something not meant to be bowed, scraping, rubbing, preparing, striking, blowing across rather than into. Voice and breath too.
Augmented instruments — a standing interest. If you play something you have extended yourself, with sensors, pickups, preparations, feedback, motors, a hybrid of acoustic body and electronics, I would particularly like to hear from you. Not only for a recording: if you want to work on something together around an augmented instrument, say so and we will talk. That is an open door rather than a one-off ask.
There is no notation to follow. You do not need to read music, and you do not need a brief. If a blank page is unhelpful I will happily write you one — a page of shapes and words, made for whatever you actually have to hand — but it is an offer, not a requirement. Most of the time the better move is that you send what you already do, and I work from that.
Take part → Already recorded? Send files →
Where it stands
Some pieces exist already. I made them to find out whether the idea holds, and their corpora came from Creative Commons recordings on Freesound — a dripping tap, bowed crotales, a prepared piano somebody else recorded years ago. They work. They are also, in a way I cannot argue myself out of, made from other people's leftovers.
That is the part I want to replace. A corpus built from people who agreed to be in it is a different object from one built out of search results. These tracks are the proof of concept; the album should be the real thing.
Listen to the studies on SoundCloud →
For the curious
The Max patcher I use, exactly as it runs. It needs Max 9 and the Somax 2.7 package installed — without them it will open but do nothing useful.
This is one of several. Others are in progress for this album — different routings, different weightings, some built around a particular kind of source rather than a general corpus — and as they settle I will put them here too. If you are the sort of person who would rather read the patch than the description, tell me and I will send whatever is current, finished or not.
Ask me anything about how it works. I would rather explain it than have it look like a trick.