Listening LoopA field notebook for your ears

Outside is closer than you think

Take a walk.
Bring back a sound.

Three places to pause. Six seconds to listen.
A little less scrolling, a little more noticing.

Make your first loop
Before you go

Pack a tiny audio model.

Download once, then take it outside. About 19 MB including the app. No account, API key or recording uploads.

Your microphone stays off until you press Listen.

Your listening loop

Find three places to pause.

0 of 3 stops
  1. 01Start close to home.
  2. 02Try a different soundscape.
  3. 03Listen again on the way back.

Pause somewhere comfortable. Put the phone down and listen first.

Only your notes and sound labels are saved. Audio is discarded after analysis.

A suggestion, not the last word

Your ears get the final say.

The model recognises broad sounds: birds, wind, traffic, voices. It can miss things or get them wrong. Add what you actually heard to your journal.

This isn’t a species identifier, a noise meter or a safe-route guide.

Try it before heading out

A small song, from a spruce tree.

Run the model on a public-domain birdsong recording. It won’t be added to your walk.

Example: Oona Räisänen, Finland, 2007. Not a recording from your walk.

How the open model works · privacy · credits

Google’s YAMNet (Apache 2.0), running through TensorFlow.js, recognises 521 audio-event labels. We average model scores over short audio frames. Scores are not calibrated probabilities; sound groups and uncertainty cutoffs are simple, editable heuristics.

All inference happens in your browser. This app has no analytics, location access or audio upload endpoint. Loading the site downloads static files from its host. Recordings are held in memory briefly and discarded after analysis. Exported journals contain labels and your written notes, never audio.

After successful preparation, the app and model are cached for offline use. Browser storage can be evicted; check the readiness message before heading out.

Built by Kanishq Sharma with disclosed AI-assisted implementation and testing for Hacktoberfest: Touch Grass. Model card · TensorFlow.js · Third-party notices