Machine learning detection of cetacean tonal calls without human annotations

Abstract

Deep learning has shown great potential in species-level detection and classification of marine mammal vocalizations. Recent studies suggest that training a neural network that performs well on unseen data often requires a large amount of analyst-annotated data, which are not only very expensive but also time-consuming to obtain for marine mammal species. We propose to develop new methods to learn such deep models from raw acoustic data with minimal human supervision in the loop. The proposed research will lead to a portable end-to-end whistle and moan detector capable of dealing with a wide variety of species, marine settings, weather conditions, etc. An option year provides for developing a classifier to extend the system to concurrently predict species from the detected tonals.

Document Details

Document Type
DoD Grant Award
Publication Date
Jun 09, 2021
Source ID
N000142112567

Entities

People

  • Marie A. Roch

Organizations

  • Office of Naval Research
  • Salk Institute for Biological Studies
  • United States Navy

Tags

Readers

  • Marine Mammal Biology
  • Neural Network Machine Learning.

Technology Areas

  • AI & ML
  • AI & ML - Neural Networks