Case study / Museum concept / 2026

Can you teach children to doubt a voice without making the experience feel scary?

This concept translates thesis work on audio deepfake detection into a museum-friendly format for NEMO: short, visual, touch-first stations that explain how fake audio works, why compression matters, and where both people and AI can get fooled.

2026conceptAI literacyAudio deepfakesMuseum UX

Audience

Children ages 8-14, museum educators, and general visitors

Role

Concept design, interaction design, research translation

Core question

How do you explain fake audio and model fragility without turning the experience into either a lecture or a gimmick?

Why this matters

Fake audio is persuasive precisely because it sounds ordinary.

The exhibit concept stays grounded in media kids already encounter: voice notes, compressed clips, distorted calls, and the fuzzy gap between what sounds convincing and what is actually true.

01

Deepfakes are often discussed visually, but fake audio is arguably more intimate. It travels through voice notes, calls, livestreams, podcasts, and all the small formats that already feel personal.

02

Children do not need a fear-based lecture. They need a way to practice doubt, see how easily certainty collapses, and leave with sharper instincts about manipulated media.

03

Compression is the bridge between the museum concept and the thesis. It turns a purely academic detector story into a real-world question about how audio behaves once it leaves the lab.

Experience design

Three stations, each short enough for a museum floor and rich enough to carry the thesis.

The core exhibit keeps every interaction tactile and low-friction, then layers in the research story only once curiosity is already there.

01

Echt or Fake?

Children hear a short clip, make a fast guess, and immediately see whether their intuition matched the answer.

Interaction

Listen, vote, reveal

Takeaway

Fake audio is not always obvious. Confidence and correctness do not always line up.

02

Break the Sound

Visitors add compression, noisy-room artifacts, and phone-like quality to learn how everyday audio conditions hide clues.

Interaction

Drag sliders, hear the difference

Takeaway

Bad audio quality changes what humans hear and what AI can reliably detect.

03

Beat the Machine

Instead of cloning a voice, visitors try to confuse the detector with realistic distortions and learn why robustness matters.

Interaction

Push the model off balance

Takeaway

The problem is not only fake speech generation. It is also how fragile models can become once the recording conditions shift.

Project facts

Status

Concept

Audience

Ages 8-14

Format

Museum installation

Focus

Audio deepfakes

Interactive teaser

A front-end-only sketch of the first station.

The scaffold ships a mock interaction for “Echt or Fake?” so the page already feels alive without depending on audio assets, spectrogram images, or inference code.

Try the interactive demo

Echt or Fake?

Clip A / Studio apology
1 / 3

This mock version keeps the structure of the first station. The clips and clue view are placeholders, but the interaction flow is already close to the intended museum behavior.

A short, polished apology recorded with a very clean microphone.

Signal sketch

Make a choice to reveal the explanation and move to the next clip.

Thesis translation

The research story in plain language.

This is where the page shifts from concept art into portfolio case study: less “AI magic,” more explanation of why real-world audio quality changes what both people and models can hear.

Idea 01

The detector does not understand speech the way people do. It looks for patterns in the sound signal that correlate with real or generated audio.

Idea 02

That works better on clean recordings than on messy, compressed, everyday clips. Voice notes, cheap microphones, and noisy rooms can hide the exact clues the detector wants to use.

Idea 03

My thesis work focuses on making detection more robust when the recording conditions change, so the model is not only good in the lab but also more reliable in the real world.

Museum constraints

The exhibit only works if it survives real museum conditions.

Children move quickly, rooms are noisy, screens are shared, and the story needs to remain understandable at first glance. The design is shaped around those constraints instead of treating them as polish issues.

Constraint 01

The main interaction needs to work in under two minutes with almost no reading.

Constraint 02

The exhibit has to survive noisy shared spaces, not quiet headphones-only conditions.

Constraint 03

Ages 8-14 is a wide range, so every station needs a simple layer and a curiosity layer.

Constraint 04

No live voice cloning in the first museum concept. Agency comes from exploration, not impersonation.

Constraint 05

Every screen should reset cleanly for the next visitor without staff intervention.

Constraint 06

The experience should be ready to expand into Dutch and English without rewiring the component tree.

Next steps

What would turn this from a strong case-study scaffold into a museum-ready prototype.

The first pass proves structure, tone, and interaction direction. The next passes would replace the mock layer with real media, bilingual copy, and exhibit-specific UX details.

Step 01

Replace the mocked clue layer with real, permission-safe media and a first working playback flow.

Step 02

Add bilingual copy, exhibit reset states, and guidance for museum facilitators or floor staff.

Step 03

Prototype the compression station next so the portfolio page evolves from a case study into a usable demo shell.