Name the sounds

Where these came from. Not a detector — the loudest distinct moments in the isolated noise channel, once the voice is separated out. That deliberately catches a mix, and whatever turns out to be nothing is as useful as whatever turns out to be a ruffle, because negatives are exactly what the classifier lacks.

The orange button is the sound on its own, with the speech removed and the level lifted so it is audible. The green one is the same moment in the full mix, for when the isolated version is ambiguous. Each clip starts one second early.

Worth knowing before you spend the time. The classifier scored 0.50 across episodes — a coin flip. Four episodes of labels might fix that or might show the same problem with more data; I cannot promise it works. But the labels are worth having either way: every sound effect you mark is something no automatic approach should ever touch, and that is worth knowing however we end up detecting.