Brief

Convey is a design platform founded by Simple Flair in 2023, running for a week during Milan Design Week as an international event dedicated to contemporary design. The 2026 edition took over a five-floor residential building near Torre Velasca. Each apartment and room in the building was given over to a brand or designer presenting their work, so the common areas were the only spaces that connected the programs together. The event housed more than twenty international brands, among them USM, Completed Works, Cimento, Max Design, Terraformae, and Western Acoustics, alongside objects and spaces by designers including Patricia Urquiola, Stefan Diez, Panter & Tourron, and BIG-GAME, with Rooms Studio as guest designer. Bizzi & Partners served as main partner, with additional support from Levi's and Fever-Tree. The rooftop terrace was the event's social centerpiece.

The initial conversation, led on our side by Giulia Spitaleri, was open: Convey's team wanted sound in the common areas. No further specification. We visited the building and noticed immediately that the common areas people would actually spend time in weren't the landings or the terrace but the stairs themselves. The building's vertical circulation was going to be the site where visitors slowed down, got stuck behind each other, caught their breath between floors. It is also, in a residential building, where neighbors meet by accident: the informal social circuitry of the condominio, and the space where most of its unscripted encounters happen. Stairwells are notorious listening environments, full of long reverbs and hard surfaces, but those same qualities can make them among the most sonically memorable spaces in a building. We proposed the stairwell as the installation site, with all the acoustic difficulty that implied, and Convey agreed.

From kickoff to installation we had just over a month, a compressed window for a project covering concept development, contribution intake, composition, processing pipeline, sound system design, and install. Several of the decisions documented below were shaped by that constraint.

Contribution system

Before writing any music, we built the contribution system. It had to be friction-free to get a usable volume of voices in the time available, and Thomas built the first working version over a weekend before any notes were written.

Contributors visited a mobile-first web app that presented a scale, one note at a time. Each note could be previewed as a reference tone, then recorded by the contributor. The instructions asked them to "hum or sing," and we kept the prompt deliberately loose. In practice, contributors did a wide range of things: sustained notes, syllables, hummed tones, spoken fragments, and in at least one case, beatboxing. We decided early on to preserve the full range of interpretation rather than filter it.

The scale was heptatonic. We defaulted to octave 4, with an option to record additional octaves for contributors comfortable doing so. We tested varying the requested octave dynamically, but observed that non-singers reliably defaulted to their own natural register regardless of what we asked for. We stopped trying to direct the octave and started measuring it in post instead.

An early version of the interface included a real-time tuner. We removed it after observing that it discouraged contributors more than it helped them. Pitch correction moved to the backend, with a deliberate decision to nudge rather than fix: we preserved most of the pitch variation within each recording, correcting only the overall center of the note.

The stack was a Next.js frontend, a Cloudflare Worker to handle ingestion and upload, and a Linux server at the studio for processing. Device compatibility took continuous iteration throughout the contribution window. Several Android models refused getUserMedia calls for WebM audio; others applied limiters or compressors that altered the raw recordings. We shipped small fixes throughout the intake window and preserved every submission in its original form, so any corrections could be made in post rather than lost.

Composition

Composition began with around thirty voices. A script ran every morning to rebuild the sampler from the latest contribution pool, which meant the instrument under our hands mutated daily. Some mornings the sampler was more out of tune; some mornings a strong singer contributed and the composition abruptly turned beautiful.

The piece was authored in Interval, an internal programming language we've developed for compositions with evolving harmonic and rhythmic structure.

It allowed quick iteration of ideas through an algorithm-like approach to music. An example:

// Interval 0.7
// Condominio — sample eight-speaker routing over eight bars.

@ppq 480
@bpm 50
@ts 4/4
@bars 8
@seed 217
@title "Condominio"

// A chromatic descent in a minor context — modal centres slide
// under a mostly static vertical field.
@scale root=C mode=minor * 4 | root=B mode=minor * 2 | root=A mode=phrygian

@harmony
i | i | iv | iv
i | bVII | bVI | i

// ---- patterns ---------------------------------------------------

@pattern rest unit=1/1
  .

// Floor 4 · Lower — upper cluster, enters late, breathes
@pattern f4l unit=1/2
  .
  {^2+^3+^4, ^2+^4, ^3+^4}[vel:58 prob:0.7]

// Floor 3 · Upper — the most active middle voice, evolves
@pattern f3u unit=1/4
  ^6+^7+^1[vel:72]
  ~
  {^6+^1, ^7+^1, ^6+^7+^1}[vel:66]
  ~

// Floor 3 · Lower — sustained third/fourth interval with agogic lean
@pattern f3l unit=1/2
  ^3+^4[vel:70]
  ~

// Floor 2 · Upper — doubled with F3·Upper an octave down, thinner
@pattern f2u unit=1/4
  ^7+^1[vel:64 prob:0.85]
  ~
  ~
  {^1, ^7+^1}[vel:60]

// Floor 2 · Lower — bass cluster foundation, steady
@pattern f2l unit=1/1
  ^4+^5+^6[vel:80]

// Floor 1 · Upper — the deepest sustained band
@pattern f1u unit=1/1
  ^1+^2+^3[vel:78]

// ---- tracks -----------------------------------------------------

@track floor_4_upper ch=1 oct=5
  play: rest

@track floor_4_lower ch=2 oct=5 gate=0.95
  play: f4l -> vary(0.4) -> swell(0.6, arch) -> humanize(25ms, 0.4)

@track floor_3_upper ch=3 oct=4 gate=0.9
  play: f3u -> evolve(0.12) -> humanize(15ms, 0.3) -> vel_curve(wave=sine, min=55, max=82, repeat=2)

@track floor_3_lower ch=4 oct=4 gate=0.92
  play: f3l -> agogic(2, 5, 8) -> humanize(20ms, 0.25)

@track floor_2_upper ch=5 oct=3 gate=0.88
  play: f2u -> vary(0.3) -> humanize(18ms, 0.35)

@track floor_2_lower ch=6 oct=3 gate=0.96
  play: f2l -> swell(0.5, ease_in_out) -> humanize(30ms, 0.3)

@track floor_1_upper ch=7 oct=3 gate=0.96
  play: f1u -> humanize(30ms, 0.3)

@track floor_1_lower ch=8 oct=2
  play: rest

The original intention was generative and live: contributions would flow continuously into the sampler, round-robin playback would differentiate each pass through the composition, and visitors would be able to record their voice on-site and hear themselves enter the piece in close to real time.

A nasty bug, triggered by an unusual submission format, convinced us during the final week to reduce the live system's error surface. We also moved from Ableton Live to Steinberg Nuendo for faster and more flexible stem export. We exported synchronized stems for each speaker instead and played them back from a fixed file set. The daily rebuild of the sampler preserved most of what we wanted from the live system: every time we re-exported, the composition altered because each note pulled a different contributor. Playback became reliable enough to run unsupervised.

The initial musical direction was closer to minimalism than to what the piece became. Vocal arpeggios running up and down the stairwell, a sparse layer of activity distributed by register. What emerged was a hybrid between that and a contemporary choral idiom, with Ligeti as the primary anchor in the cluster-writing lineage and Holly Herndon's PROTO as a conceptual reference more than a sonic one.

The five-floor register logic survived the production process but softened. The original plan was clean range separation by floor. What we ended up with is closer to a pyramidal distribution: each floor foregrounds its dominant register, but all layers carry through the whole system via controlled send levels. Lower floors lean on the tenor material while keeping mid and high content present; upper floors sit in the soprano register while retaining a subtle presence of everything below. The result reads as more organic and, in the finished piece, more faithful to the way voices actually carry in a stairwell.

Tuning and acoustics

The 2–3 second reverb became a compositional asset: it extends the tails of the vocal material and glues the layers together across floors. The speaker directivity carries the intelligibility, the room carries the blend. No acoustic treatment was introduced; the small drivers' limited low-end extension handled most of the usual stairwell problems on its own. Per-floor EQ was applied by ear during installation, primarily cutting low frequencies and more aggressively on the lower floors where the tenor material concentrates energy in the low-mid range.

Contribution and credit

Condominio is built from fifty-two voices as of the installation's first day. Contributors were gathered publicly through open outreach, ranging from close collaborators of the studio to people at one or two removes through Convey's and Hidden Mountain's networks.

Every contributor is named. The mapping.json file produced by the build pipeline pairs every note in the instrument with the contributor whose voice plays it, and a visualizer we built surfaces the contributor currently playing in real time, used in press and social material about the piece. Printed credits distributed at the installation list every participant in full.

We avoided heavy on-site visual treatment. Stairwells are circulation spaces, not viewing rooms, and the last thing we wanted was to encourage crowds to stop mid-ascent to read wall text. The participatory side of the piece reads most clearly in the documentation, communications, and visualizer, not in the installation itself, a tradeoff we'd revisit in a future iteration.