Sofilab - Interactive Sound Design

02/06/2026

Most product sound systems work like this: trigger event, play file. Door opens: play chime. Error occurs: play alert. Battery low: play warning. A library of audio files mapped to a list of events.

We build something different: a semantic operating system for sound. Instead of mapping triggers to files, it maps meaning to generated audio in real time.

The architecture has four stages. Context (vehicle dynamics, environment, cabin state, user state) flows into a Semantic Layer that interprets what's happening and what it means. That interpretation feeds into our future CORPUS Reef, a real-time generative sound engine. The output is continuous and adaptive: soundscapes, interaction sounds, sonic brand identity, and adaptive silence.

The shift is fundamental. You stop tuning individual sounds and start designing meaning. Instead of asking "what should the parking chime sound like?" you ask "what should approaching an obstacle feel like?" The system handles the translation from meaning to sound, adapting to context every moment.

This is what makes sound scalable for complex products. An autonomous vehicle has thousands of possible states. You can't hand-craft a sound for each one. But you can design the semantic rules that govern how all of them should feel.

More on the architecture: sofilab.art/

28/05/2026

Classification flattens, description holds detail. ran well-known tracks past the market-leading taggers, and the results show the cost of the closed list: a Yoruba Afrobeat track read as Latin, a fado read as Klezmer. The CORPUS pipeline describes instead of filing.

Carousel below.

26/05/2026

The CORPUS contribution app went public today. For years at Sofilab we kept hitting the same wall: the generative models worth putting into real products need training data with clean provenance, and that data did not exist to buy. CORPUS is the infrastructure we are building so that it does, and the app is where musicians start contributing to it on their own terms.

Link below.

Most upload platforms take your music the instant you drop it in. Ingested, indexed, put to use, often before you have finished reading the terms.

The CORPUS contribution app opens the other way around. As a visitor you can upload anything, a rough mix, a record you love, a track you are unsure about, and none of it enters the corpus. What comes back is how our engine reads the music: a deep, written description, the genres it actually serves, the contexts where it would belong.

The app is public today. No invite, no waitlist. After a long run of "almost there," the project is open to anyone who wants in. You can walk in as a visitor and look around with nothing at stake, or step further and contribute.

If you contribute, you are joining a phase where the things still left to settle get settled with the people who joined in the room: the commercial license, the rules of contributor ownership, how scoring works, the questions we have not answered yet. The public beta runs under a provisional license that grants us no commercial rights, on purpose: this is the phase to work the structure out together, before any commercial use. Your music enters a commercially usable training set later, only if you agree to that step.

This is what we have been building toward. From here on it is yours to use, and yours to shape.

Go in. Upload something. Tell us where the description is wrong, and where the project itself is wrong.

Link in the first comment.

04/05/2026

Sound on.

When HARMAN Automotive asked us to design sounds for AutoOne, the brief was contained: demo music plus start-up and shut-down sounds for HarmanKardon, for the CES stand, mixed and rendered in Dolby Atmos. By the time Jörg Hüttner and Mathis Nitschke delivered, the project had grown into something larger.

A start-up sound is the first impression a system makes. So before any sound was designed, we started with the brand. What does Harman Kardon actually stand for, sonically? What's the DNA that should run through every touchpoint?

​The answers shaped every individual sound and became an impulse for a broader conversation about the brand's sonic identity.

​What started as a sound design brief evolved into a redefinition of how Harman Kardon presents itself inside a vehicle.

Presented at CES 2025.

Thank you Sabrina Ceylan, Tanaponn Kulsedzeranee, Philipp Siebourg and team for the trust and the creative collaboration.

29/04/2026

A soundwalk that works anywhere in the world.
​
Most soundwalks are tied to a specific location: pre-composed content triggered by GPS coordinates. Walk outside the route and the experience breaks. City Songs takes the opposite approach.
​
Instead of composing for a fixed path, we built a context classification pipeline. The app reads the listener's real-world environment (weather, sun position, temperature, environment type) using custom parsers for OpenStreetMap data. It classifies the context and generates a site-specific sonic experience in real time.
​
Walk through a park in Tokyo, a street in Lagos, a harbor in Bergen: each produces a different experience, because the parameters are different. But the musical logic is the same.
​
The system also uses GPT-4 for location-specific poetry, weaving textual fragments into the generative audio layer based on the cultural and geographic context of the listener's position.
​
Why this matters beyond art: the context classification pipeline we built for City Songs is the direct precursor to our adaptive automotive sound systems. The same architecture (sense context, classify, generate) now runs in vehicles. From soundwalks to product sound, the underlying problem is identical: make sound that responds to where you are and what's happening around you.

An experimental project funded by FilmFernsehFond Bayern.
​

27/04/2026

Before a sound becomes a melody or a product tone, someone has to create the raw material: the timbre, the texture, the sonic building block. That's what Jörg Hüttner does. He's a sound creator and synthesizer programmer at the deepest level.
​
Jörg programs synthesizers for Native Instruments, Arturia and other leading manufacturers, and creates sounds for major film and TV productions. His synth programming and sound design credits include Apple TV's "Silo" (with Atli Örvarsson), Netflix's "Army of Thieves" (with Hans Zimmer), and Roland Emmerich's "Moonfall." When these productions need a specific sonic palette, Jörg builds it.
​
At Sofilab, Jörg is the person who creates the fundamental sound material for our projects: the timbres, the textures, the sonic DNA that everything else is built on. Whether it's a startup sound for Harman Kardon or a feedback tone for a robotic system, the character of the sound begins with the synthesis.
​
20+ years across film, gaming, and industrial sound. A sound designer who creates sounds from scratch, rather than a composer who uses them. The difference matters when every millisecond and every harmonic detail counts.
​
Hear his work on our website → https://sofilab.art

23/04/2026

Nearly every Sofilab project of the last decade has hit the same wall: product sound that has to read a situation and keep reading, with no one in the loop at the moment of generation. That is the territory Mathis' Journal article describes, and why Corpus.Music exists. The song-generator debate is happening several doors down.

The public AI music debate has narrowed to a single question: who gets paid when a machine generates a song.

The question is legitimate. It operates inside a frame that is already being outgrown.

In cars, hospitals, games, rehabilitation, the finished song is not the product. What is needed is sound that reads a situation and keeps responding to it, in real time, on device. These markets are regulated, context-specific, machine-to-machine. Anything built around the prompt-in, song-out flow is architecturally incompatible with them.

The major-label settlements with Suno and Udio addressed a legal exposure. The verification problem underneath is untouched: once a model is trained, no one can audit what it was actually trained on. "Trained only on licensed music" is a contractual statement. The architecture cannot prove it either way.

We wrote a new Journal article about this, because the infrastructure these adjacent markets need (a semantic layer, real-time generation on device, and an auditable training corpus) is what we are building at CORPUS.

What is arriving is something older: sound that responds to the situation it is in.

Link to the article in comment.

22/04/2026

An operating room is one of the most acoustically demanding environments in professional work. Patient monitors beep. Ventilators cycle. Alarms overlap. Every device competes for the same narrow band of attention.

Into this environment, a robotic imaging system added its own audio cues; they were indistinguishable from the existing monitor sounds. Same frequency range, same temporal patterns, same urgency coding. The result was cognitive overload.

We designed a non-verbal communication system for the robot. Distinct sound signatures for different movement types. Graduated audio cues for radiation intensity levels. Clear state indicators that separated the robot's communication from everything else in the room.

The design principle was reduction, not addition. Every sound had to earn its place. If a visual indicator was sufficient, the sound was removed. If a sound was necessary, it had to be instantly distinguishable from patient monitoring equipment.

The goal was simple: allow surgeons to keep their attention on the patient. Reduce cognitive load instead of adding to it. Let the robot communicate its state without demanding the surgeon's eyes leave the operating field.

Sound design in medical robotics is about making critical information available without stealing attention from what matters most.

20/04/2026

CORPUS Journal examines how AI is reshaping music, creativity, and cultural systems.

11/09/2022

Experience the famous orchestral suite "The Planets" by Gustav Holst as an interactive audio walk with smartphone and headphones.