Who's singing? Automatic bird sound recognition with machine learning
summary
System learns species from 10‑second mobile recordings, sending audio to a cloud service that returns a ranked list and stores data as open archives. Feature pipeline uses spectrograms, Librosa for audio I/O and preprocessing and learned bases from spherical k-means rather than fixed templates. A public challenge yields up to about 90% scores and the research becomes a working product, Warblr, with roughly 5–6k paying users; offline use is a known limitation.
AI-written summary · about
chapters (11)
- 0:00Bird sound recognition intro
- 1:36Machine Listening Lab research
- 4:05Why specialized recognition
- 9:28Audio processing tools
- 10:13Research to Warbler app
- 16:12App architecture success
- 19:48Ecological monitoring datasets
- 23:02Bird detection vs classification
- 24:50DCASE challenge details
- 27:18Real-world applications
- 29:46Q&A
related talks
15mBirdNET: AI listening for life
11mAI and Bioacoustics: Tracking Biodiversity Progress with BirdNET
39mOverview and Marine Bioacoustics and Perch Model
23mBirding with Python and Machine Learning
53mSelf-Supervised Learning in Transformer Models for Birdsong
29mBioacoustics and Machine Learning as Key Tools in Conservation
25mThe Sound of Life
40mBuilding Features and Machine Learning Models for Audio Signals