DCCEA study of sperm whale codas
Bioacoustics / Representation learning

What can a
coda tell us?

Exploring rhythm, spectral information, and identity in sperm whale communication.

João Victor Quintanilha · Emma Virnelli · Shuo Zhang
Tufts University

Explore the study →Read the paperSee the presentationView the poster
One coda, two input representations
RhythmInter-click intervals
SpectrumFull-coda log-mel input

Schematic illustration · listen to real recordings

1,383clean codas
3social units
22coda types
11eligible individual identities
The research question

Separate inputs.
Task-dependent results.

Does encoding timing and spectral information separately improve what a model can learn from sperm whale codas? DCCE tests this idea across social unit, coda type, and individual identification.

Simple timing features remain strong. Benefits depend on supervision, and performance falls under recording-date holdouts. The study supports task-dependent combinations of information, without establishing independent biological channels or general architectural superiority.

Read the full study
01From recordings to a labeled dataset02How the dual-input model works03How a fair comparison is defined04What the experiments support05What changes across recording contexts06Inspect the evidence
Three things to take from the study

Timing is a strong reference

Raw ICI gives 0.987 ± 0.009 random type macro-F1. Annotated timing should be explicit in comparisons with waveform-only models.

Compare representations

Branch gains can be lost in fusion

Adding the identity objective improves the spectral branch, but does not improve the mean fused-ID result in the main configuration.

Inspect the ablations

Context limits generalization

Grouped and year-disjoint evaluations expose weaknesses hidden by a single random split. Different class support also changes what is being measured.

Read the limitations