Deep Learning and Domain Transfer for Orca Vocalization Detection

Abstract

In this paper, we study the difficulties of domain transfer when training deep learning models, on a specific task that is orca vocalization detection. Deep learning appears to be an answer to many sound recognition tasks in human speech analysis as well as in bioacoustics. This method allows to learn from large amounts of data, and find the best scoring way to discriminate between classes (e.g. orca vocalization and other sounds). However, to learn the perfect data representation and discrimination boundaries, all possible data configurations need to be processed. This causes problems when those configurations are ever changing (e.g. in our experiment, a change in the recording system happened to considerably disturb our previously well performing model). We thus explore approaches to compensate on the difficulties faced with domain transfer, with two convolutionnal neural networks (CNN) architectures, one that works in the time-frequency domain, and one that works directly on the time domain.

Intra-Group Orca Call Rate Modulation Estimation Using Compact Four Hydrophones Array

“…To accomplish this, we analyzed three different natural scenes or observations. Previous shore-based fixed hydrophone systems enabled the detection and tracking of orcas from the same population (Grebner, 2009;Bergler et al, 2019;Poupard et al, 2019a;Best et al, 2020) but did not succeed in associating individuals with calls or determine precise individual pattern variations of communication between groups. Tracking of a few individual orcas in another population, without visual checking, has been realized in a complex pilot study using 14 hydrophones deployed in three compact arrays (Gassmann et al, 2013).…”

Evidences of Intra-Group Orca Call Rate Modulation Using A Small-Aperture Four Hydrophone Array

“…Our shore based study looked at orca communication in the wild at a unique level of precision. Other shore based fixed hydrophone systems enabled the detection and tracking of orcas from the same population [36][37][38] but did not succeed in associating individuals with calls, or determine precise individual pattern variations of communication between groups. This work does not pretend to provide all explanations for the type of calls, but rather the CR.…”

Multi-Attentive Detection of the Spider Monkey Whinny in the (Actual) Wild

“…Reshape embedding (9,8192) functions -i. e., multi-head attention -in the context of audio MIL classification. Assuming K attention heads, the aggregated bag-level embedding per head is calculated as follows:…”

“…Automated methods for recording and analysing bioacoustic data hold the promise for unprecedented scalability in wildlife monitoring, with the purpose of preservation through a global biodiversity crisis [1]. This has enabled biologists and engineers to perform machine learning studies on bioacoustics across a large taxonomic range, such as primates [2,3] or other terrestrial [4,5] or marine mammals [6,7,8,9,10], birds [11,12,13,14,15], as well as amphibians [14], in applications like call detection for verifying presence or estimating density [6,2,4], discerning between calls of different species [14,15], as well as different call types of a particular animal [5,8].…”