The Duck Pond synthesizes every quack live, and you can throw the ducks peas

A developer who goes by celphase built The Duck Pond, a web page with a small flock of mallards on it: three hens, two drakes and two broods of ducklings. There isn’t a single recording in it. Every quack is synthesized as it happens and heard from where you stand on the bank, and one button throws them a handful of peas.

The ducks have names and routines. According to the project’s GitHub page, Mabel, Clover and Hazel are the hens, Rufus escorts Clover and Bramble is the lone drake, and they spend about 40% of their time feeding, 30% resting and 15% preening. Throw peas and the nearest ducks notice first and call out, then race from pea to pea. The pond’s code caps the peas at 120. A duckling that strays more than 4 meters from its mother stops and peeps until she calls it back.

The sound is the wild part. The README says the quacks model the duck’s syrinx and vocal tract, with the formants following the bill as it opens on “qu-” and closes on “-ck.” celphase wrote that it was made with significant LLM use (they’d credit Claude as a co-contributor if they could), running the LLM in a loop to bring the output’s spectrogram closer to a real recorded sample. They added that a classification model marked the synthesis as more duck than the real duck sample.

If you try it, I’d put on headphones and turn on its headphone mode, since the README says each ear gets its own delay from a model of a human head. It’s free, under the MIT license, and the code is on GitHub.

Bring peas.