Classifying bioacoustic data without individual call annotations using temporal convolutional networks and feature
Laia Garrobé Fonollosa1, Douglas Gillespie1, Lina Stankovic2
1Sea Mammal Research Unit, School of Biology, University of St. Andrews, KY16 9TH, St. Andrews, Scotland.
The Journal of the Acoustical Society of America
|August 7, 2026
Summary
This study introduces a new framework for analyzing bioacoustic data using temporal convolutional networks (TCNs), overcoming limitations of weak labels in passive acoustic monitoring (PAM). The approach effectively classifies species, like sperm whales, with high accuracy, comparable to expert agreement.
Area of Science:
- Marine bioacoustics
- Computational ecology
- Machine learning for bioacoustics
Background:
- Passive acoustic monitoring (PAM) generates vast bioacoustic datasets.
- Detailed auditing and manual labeling of PAM data are often impractical, leading to weak annotations.
- Weak annotations hinder the effective analysis of complex temporal patterns in long audio segments.
Purpose of the Study:
- To propose a framework for standardizing datasets, extracting features, and classifying complex temporal patterns in bioacoustic data using temporal convolutional networks (TCNs).
- To eliminate the need for heuristic decision rules or time-consuming strong labels in bioacoustic analysis.
- To demonstrate the framework's effectiveness using sperm whale click trains as a case study.
Main Methods:
- Dataset standardization, feature extraction (variational autoencoders and hand-picked features), and classification using temporal convolutional networks (TCNs).
- Case study using sperm whale (Physeter macrocephalus) click trains from diverse recording conditions.
- Comparison of feature extraction methods: variational autoencoders (VAEs) versus traditional hand-picked features.
Main Results:
- TCN classifiers achieved recall rates exceeding 0.83 at a 0.13 false positive rate, comparable to expert annotator agreement.
- Both VAE-based and hand-picked feature extraction methods yielded similar classification performance.
- VAE-based classifiers demonstrated more stable performance across varied datasets and recording conditions.
Conclusions:
- The proposed framework effectively leverages existing weakly annotated bioacoustic datasets for training automatic classification models.
- This approach overcomes the limitations imposed by weak labels in large-scale bioacoustic data analysis.
- The method offers a viable solution for efficient and accurate analysis of marine mammal vocalizations and other bioacoustic signals.