Playground

Walk through real soundscapes, rescue speech from noise, and find out when your eyes disagree with the metric.

Three toys about how people hear and see. Everything runs in your browser and nothing you do is recorded or sent anywhere. Headphones make the first one much better.

Turn your volume down before you press play.

Walk through a soundscape

These are real field recordings placed around you in 3D. Drag any sound, or drag yourself (the white head) and walk. Things happen on their own: a dog, a siren, a passing tram. No two minutes sound alike.

Scroll over the scene to turn your head. With the scene focused, arrow keys walk and Q and E turn.

Make something happen

In this scene

Add a sound from anywhere

How it feels

A playful estimate from what is in your mix, using the axes of ISO 12913-3. It is not a measurement.

Credits for the recordings

Catch the sentence

Someone is talking in a noisy place. Can you catch what they said? Then switch on enhancement and find out what the machine heard. The twist comes from my EMNLP 2026 paper: speech that sounds better is not always speech that is understood better.

Enhancement was run offline with DeepFilterNet and MetricGAN+. Transcripts are from Whisper small.en, PESQ and STOI are measured against the clean speech. Speech is from LibriSpeech (CC BY 4.0). Read the paper · About the project

You vs the metric

Flip enhancement on and off, then keep the repair you like. Only after you vote do you see what the numbers preferred, and what the photo really looked like.

Repairs were made offline with Real-ESRGAN, SCUNet and FBCNN. Scores are measured against the original. More on image quality metrics

Credits for the photos

Other things to click