A microphone at a moonlit wetland field station, with turquoise sound ripples crossing the water

Private acoustic field kit

See what a tiny
classifier hears.

Collect three sounds, reveal their visible fingerprints, then challenge a simple nearest-neighbor model—without sending a recording anywhere.

Open the field kit
  • Local only
  • Works offline
  • Explainable

01 Moonlit traces turn an invisible signal into something you can compare.

Your local experiment

The field kit

Make each sound a few times. The model only compares the examples you give it.

01

Name three sounds

Choose actions or objects—not people or traits.

02

Listen & collect

Select a label, then record a 1–4 second example.

Ready to listen 00:00.0

The live waveform will appear here while the microphone is listening.

No microphone permission requested yet.

Visible evidence

Inspect a sound

Record or select a specimen to see the evidence the baseline compares.

No sound under the lens yet

Your first recording will reveal a waveform, frequency trail, and MFCC-like shape.

Record a first sound

Transparent baseline

Follow the classifier’s guess

It finds the three stored examples with the most similar spectral fingerprints. Nearest votes win.

Ready for a mystery sound?

Collect one example for each label, select Test sound, then record something new.

Local collection

Your specimens

0 recordings on this device

Take your work with you

JSON includes recordings and can be imported later. CSV contains the visible features for a spreadsheet.

No black box

What happens
under the surface

The playground deliberately uses a small, fallible method. The point is to inspect where it works—and catch where it doesn’t.

  1. 01

    Listen

    Your browser captures at most four seconds after you explicitly allow the microphone.

  2. 02

    Summarize

    Web Audio turns the clip into loudness, frequency bands, and an MFCC-like cosine summary.

  3. 03

    Compare

    A nearest-neighbor baseline measures Euclidean distance between normalized spectral fingerprints.

  4. 04

    Question

    You inspect every neighbor and mark wrong guesses. No hidden training and no cloud model.

Erase every local recording?

This will permanently remove all clips, features, and label names stored by this playground on this device.