Overview and Marine Bioacoustics and Perch Model
summary
A system uses a pre-trained embedding model to embed millions of hours (Australian Acoustic Observatory: 360 microphones, ~3M hours) and then performs search over embeddings to build classifiers in hours instead of months. Global bird models (Perch, BirdNet, SurfPerch) transfer to novel tasks: godwit call types (5 call types) reach >90% ROC‑AUC with ~20 examples; yellowhammer dialect needs more examples aligning to the 5th PCA axis; beluga demo turns ~50 hours of unlabeled Arctic audio plus NOAA archives into a working classifier in under two hours via Agile modeling and Colab notebooks.
AI-written summary · about
chapters (7)
- 0:00Bioacoustic monitoring intro
- 5:20Scaling classifiers with embeddings
- 9:10Global bird classifiers
- 13:51Linear separability in embeddings
- 19:53Bird models for whale vocalizations
- 26:05Non-human communication workshop
- 37:33Q&A
related talks
29mBioacoustics and Machine Learning as Key Tools in Conservation
40mWho's singing? Automatic bird sound recognition with machine learning
11mAI and Bioacoustics: Tracking Biodiversity Progress with BirdNET
15mBirdNET: AI listening for life
25mThe Sound of Life
22mUbicoustics: Plug-and-Play Acoustic Activity Recognition
59mFaces, Flukes, Fins, and Flanks: How Multispecies Re-ID Models are Transforming Our Approach to AI..
23mBirding with Python and Machine Learning