Identifying Species by Sound

CWS places acoustic recorders in the field to monitor species. The recorders produce enormous quantities of audio, most of it wind, rain, insects and silence. Somewhere in it are the calls that matter.

An engineering student is already building a classifier that identifies species from these recordings. It works. There is more data and more analysis than one person can get through.

The bottleneck is labelled data and validation, not model architecture. Someone has to listen to clips and confirm what is in them, and until that exists the classifier cannot be extended to new species or trusted for new sites.

Labelling audio is slow and needs someone who can tell two similar calls apart. Field recordings are noisy in ways that clean training data is not. And a classifier used in research has to be right in a way a demo does not: a false positive in a species survey is a false scientific claim.

• Branch — Electronics and CS. One of the few problems here with genuine signal processing content.
• Technical skills — Audio signal processing, machine learning for classification, dataset construction and labelling, evaluation under class imbalance, reproducible research practice.
• Other skills — Scientific writing, working inside someone else’s project, reporting negative results honestly.