# Detection of blue whale vocalisations using a temporal-domain convolutional neural network

**Authors:** Bryan Sagredo, Sonia Español-Jiménez, Felipe Tobar  
**arXiv:** [2110.02151](https://arxiv.org/abs/2110.02151)  
**Date:** 2021-10-05  
**Summary page:** https://www.summarizepaper.com/en/paper/2110.02151v1/

## Key points

- Authors present a framework for detecting blue whale vocalisations from acoustic submarine recordings
- Methodology consists of three stages: preprocessing, label-propagation, and convolutional neural network analysis
- Preprocessing stage enhances recording quality and reduces background noise interference
- Label-propagation mechanism ensures consistency of annotations for whale vocalisations
- Convolutional neural network is employed to analyze audio samples and detect blue whale vocalizations
- Network trained using 28 real-world submarine recordings, evaluated on six additional test recordings
- Results demonstrate promising performance indicators including accuracy of 85.4% and recall rate of 93.5%
- Approach able to identify potential parts of unlabelled whale calls due to incomplete calls
- Study presents an effective framework for detecting blue whale vocalizations from acoustic submarine recordings

## Summary

In their paper titled "Detection of blue whale vocalisations using a temporal-domain convolutional neural network," authors Bryan Sagredo, Sonia Español-Jiménez, and Felipe Tobar present a framework for detecting blue whale vocalisations from acoustic submarine recordings. The proposed methodology consists of three stages: preprocessing, label-propagation, and convolutional neural network analysis. During the preprocessing stage, the audio recordings are conditioned through various techniques such as normalization, filtering, and denoising. This step aims to enhance the quality of the recordings and reduce background noise interference. The next stage involves a label-propagation mechanism that ensures the consistency of annotations for whale vocalisations. This mechanism helps maintain accurate labeling throughout the dataset by propagating labels based on similarity measures between different segments of vocalizations. Finally, a convolutional neural network is employed to analyze audio samples and detect blue whale vocalizations. The network is trained using 28 real-world submarine recordings while its performance is evaluated on six additional test recordings. The results demonstrate promising performance indicators including an accuracy of 85.4% and a recall rate of 93.5%. Furthermore, even in cases where the detector did not match ground-truth labels precisely, visual inspection confirms the ability of their approach to identify potential parts of whale calls that may have been unlabelled due to incomplete calls. Overall, this study presents an effective framework for detecting blue whale vocalizations from acoustic submarine recordings. The combination of preprocessing techniques, label propagation mechanisms and convolutional neural networks contributes to accurate detection with high performance indicators.