AI finds manatee calls that a conventional sound classifier misses

In an earlier survey at Cameroon’s Lake Ossa, researchers detected African manatees in 29 of 120 listening surveys, compared with four of 120 visual scans. Both methods watched the same ten sites over twelve monthly visits. The view from above the water was leaving much of the evidence unheard.[2]

A 1 October 2026 paper used roughly 600 recording-days from 52 sites in the Republic of the Congo’s Conkouati-Douli National Park, collected during March 2024–April 2025.[1]

Knowledge borrowed from other sounds

One of the models, Dasheng, was pretrained on 272,000 hours of general audio. Its developers describe a masked-audio learning task: recover missing information from the surrounding signal. The released models turn sound into numerical representations that can subsequently be used for classification. Their initial education includes environmental sound, speech and music.[3]

That is useful preparation for a specialised recogniser. It need not learn every regularity of a waveform from its first few examples of the target animal. It can start with features learned from a much broader acoustic world, then use labels to identify which combinations matter for the new task.

The experiment fine-tuned ResNet on spectrogram images and Dasheng on audio, augmenting training examples with weighted mixtures.[1]

Think of a clear call and a separate noisy recording. Combining them at different strengths creates several versions of the same recognition problem. It cannot supply a new animal or a new habitat, but it can change how exposed the call is against its background. That distinction explains why extra training examples and extra field observations do different jobs.

Fewer labels, still a broad field sample

On a balanced test set from three withheld sites—319 calls and 311 other events—fully trained Dasheng recovered 90% of calls, versus 33% for Kaleidoscope. Dasheng’s precision was 89%.[1]

With roughly 300 labelled calls plus background examples, the two deep models retained F1 scores of 0.83–0.84. The authors note that even a 10% subset spans 46 sites and 11 months.[1]

F1 combines recall, the fraction of real calls recovered, with precision, the fraction of reported calls that are genuine. It is not the percentage of all recordings classified correctly. The balanced test also gives calls much more prominence than they may have in a long archive; the share of false alarms during archive processing needs its own assessment.

The project makes that assessment practical to examine. Its public repository provides training and prediction scripts, and points to downloadable recordings, annotations and model weights. Prediction produces a CSV file; separate evaluation scripts compare outputs with labelled events. Researchers can inspect a sequence of candidate detections as well as a headline score.[4]

Earlier fieldwork established a related boundary: listening records a manatee’s presence when it vocalises within range. A silent animal can escape detection, and a count of calls is not automatically a count of individuals. The Lake Ossa researchers therefore treated detected vocalisations as evidence of occurrence rather than a direct population census.[2]

Consider a map assembled from repeated surveys. A confirmed call can place the species in a location at a particular time. Several calls can strengthen that record without establishing several animals. Conversely, an empty detection file can mean absence, silence, an obscured signal or a recogniser’s miss. Better recognition reduces one of those ambiguities while leaving the others available for ecological investigation.

The practical gain is more usable evidence from existing recordings. Few annotations can teach an acoustic task when they draw on substantial pretraining and varied field collection.[1]

Report an error or suggest a correction