TurtlePond.us

Tools for Wildlife Observation

Turtle Pond / Heket:
[ Architecture ] [ Detections ] [ Setup ] [ Model ]

Detections

A Heket detection is more than a species prediction. It preserves the audio Heket heard, when and where it was recorded, the model’s classification, and the information needed for a human to review that classification.

Heket intentionally keeps the original evidence. A classifier can be wrong; the recording lets you decide.

Reviewing what Heket heard

Detections are presented with their predicted species, confidence, timestamp, and audio playback.

Heket separates stronger and less-certain detections to make review easier. You can listen to any retained recording and correct its label when Heket gets it wrong.

Those corrections aren’t merely cosmetic. Locally labeled recordings become training material when Heket builds future models.

Events

Events let you preserve and review a window of activity rather than an isolated clip.

If you hear a frog calling, Heket can create an event containing recordings from around that time. Events can also be created retrospectively when you know approximately when something happened.

Recordings associated with an event are protected from normal cleanup until the event is deleted.

Sharing a detection

Reviewed detections can optionally be shared outside the local Heket installation.

A detection can be contributed to the TurtlePond.us community corpus, where its labeled audio can help bootstrap other Heket installations. Selected observations can also be published to iNaturalist with their original audio evidence.

Sharing is explicit. Heket shows what information will be sent to each destination before you choose to share it.

TurtlePond.us contributions help other Hekets learn. iNaturalist sharing creates a public biodiversity observation.