CawTalker Wants to Talk Back to Crows

Field notes

The project separates acoustic classification from semantic interpretation: CawDex labels calls by type, while the remaining research pairs those labels with observations recorded at the moment of calling. Its current data constraint is unusually explicit—only 7.3 minutes of human-labeled call audio—despite a cited archive of 150,000 calls. The homepage reports 3,162 hand-reviewed segments and an alarm-call AUC of 0.97 under recording-grouped cross-validation, while presenting semantic translation as unfinished.