TRANSFERRING PRETRAINED AUDIO MODELS TO ARTIFICIAL NOISE AND ABIOTIC SOUND | |||
| Amy Van Scoyoc; UC Berkeley; avanscoyoc@berkeley.edu; Brookie Guzder-Williams, Nick Gondek, Michael Catchen, Lindsey Rich, Ciera Martinez, Carl Boettiger, Justin Brashares | |||
Autonomous recording units capture artificial and abiotic sounds alongside wildlife, and detecting these sounds is essential for quantifying human impacts or filtering noise in ecological analyses. While embeddings from bird-trained models such as BirdNET and Perch transfer well to non-avian taxa, their performance on non-biological sounds remains poorly characterized. We labeled 6,547 three-second clips from 788 field recordings and trained lightweight classifiers on frozen embeddings from six pretrained bird, general-audio, and image models to detect 14 types of artificial sound (e.g., generators, traffic, static) and abiotic sound (e.g., wind, thunder). For most sound types, accuracy approached its maximum with about 40 labeled examples per type, and custom classifiers usually matched or outperformed the models’ built-in classifiers. Accuracy depended more on sound type than on model. Our results indicate that factors such as average frequency, clip input length, and sample diversity matter more than model choice for adequately delineating machine noise and acoustically similar sound types. | |||
| |||
Speaker Bio: Dr. Amy Van Scoyoc is a postdoctoral researcher at the UC Berkeley Eric and Wendy Schmidt Center for Data Science and Environment. Trained as a field ecologist, she has spent the past decade studying how human disturbance affects wildlife, from carnivores to amphibians and birds. Her current work focuses on developing tools to make biodiversity monitoring more efficient and scalable. |