Model

Heket does not ship with a pretrained model.

Instead, each Heket installation builds its own model from labeled audio examples. This allows the model to become specialized for the frogs and toads, microphone, background sounds, and acoustic environment where that Heket actually listens.

Starting with community knowledge

A new Heket doesn’t have enough local recordings to train a useful model. TurtlePond.us provides community clip packs containing labeled recordings contributed by other Heket installations.

During initial setup, you can select packs for frogs and toads expected at your site along with relevant non-frog sounds. Heket downloads those recordings and uses them to train its first model.

Clip packs aren’t limited to initial setup. They can be added later whenever Heket needs examples of another sound.

Learning locally

Community recordings provide a starting point. They aren’t intended to permanently define the model.

As you review Heket’s detections, corrected and manually labeled recordings become local training examples. When the model is retrained, Heket uses those local recordings to better represent what sounds actually like at your site.

For each class, local examples progressively displace community examples used during training. Eventually, a well-observed class can be trained entirely from local recordings.

One local example replaces one community example.

Why not distribute a universal model?

A broad pretrained model can recognize many species across many environments, but Heket takes a different approach.

Heket is intended to become a local specialist. It can begin with recordings collected elsewhere, then increasingly learn the particular species, microphone, noise, and acoustic characteristics of one monitoring site.

The goal of TurtlePond community packs isn’t to provide perfect classification on day one. They’re scaffolding that makes it practical for a new Heket to begin learning.

Training

Training happens on the Heket installation. TurtlePond.us distributes labeled clips, not finished Heket models.

After training completes successfully, Heket automatically begins using the newly created model. If training fails, the existing model remains available.

Closing thought

Community clip packs are built from recordings that Heket operators choose to contribute to TurtlePond.us. Using community packs does not require contributing your own recordings.

Start with the community. Learn your site.

A Heket model belongs to the installation that trained it.