Individual Discrimination of Bottlenose Dolphin Signature Whistles Using Deep Learning
summary
Long-term Sarasota suction-cup recordings (38+ years, ~61,000 whistles; ~47,000 labeled signature whistles from ~295 individuals, 171 repeatedly) feed a database and an acoustic network of 12 recorders with edge detection on Raspberry Pi. Models use spectrograms with MobileNetV2 CNN and a DINO Vision Transformer backbone, augmented by ±10% pitch/time and NOAA noise at −6 to +12 dB. Benchmarks: 72 dolphins × 200 whistles and a 100-dolphin set reach F1 0.95 and up to F1 0.98 after augmentation.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 1:02Social and cognitive complexity
- 7:28Acoustic communication
- 14:02Acoustic monitoring potential
- 15:34Field site and methods
- 19:58Signature whistle database
- 22:50Whistle variability and stability
- 32:15Deep learning classification
- 44:01Classification results and future directions
- 46:12Q&A session part one
- 53:21Q&A session part two
- 58:50Q&A session part three
related talks
56mWhistle variability and social acoustic interactions in bottlenose dolphins
01h 00mLong term work and Machine Learning Approaches to DolphinGemma
01h 07mDecoding Dolphin Speech
39mOverview and Marine Bioacoustics and Perch Model
45mEvidence of social learning across symbolic cultural barriers in sperm whales
25mThe Sound of Life
53mAI as a tool for discovery in animal communication
29mBioacoustics and Machine Learning as Key Tools in Conservation