Home Publications Research Data Research Grants Equipment Inventions & Patents Awards Additional Research Activities Faculties & Institutions Research Areas

Recording-Specific Confounding in Killer Whale Vocalization Classification: A Cautionary Tale for Bioacoustic Machine Learning

Hauer C, Nöth E, Dobmeier T, Arias Vergara T, Maier A, Vester H, Bergler C (2026)

---

Publication Type: Conference contribution

Publication year: 2026

Original Authors: Christopher Hauer, Elmar Nöth, Tobias Dobmeier, Tomas Arias Vergara, Andreas Maier, Heike Vester, Christian Bergler

Pages Range: 339-342

Event location: Zadar

DOI: 10.1109/ELMAR71231.2026.11712240

Abstract

Opportunistic field recordings are widely used to study animal communication, but their uncontrolled acoustic conditions pose substantial challenges for machine learning. This study investigates whether killer whale group affiliation can be inferred from short vocalization recordings using deep learning models trained on spectrogram representations. Using 3,297 manually annotated vocalizations from 56 recording sessions, we evaluate multiple acoustic conditions, including denoised and residual-noise representations. Diagnostic experiments based on recording-session classification and session-held-out evaluation reveal that apparent group classification performance is primarily driven by recording-specific acoustic structure rather than stable group-specific vocal cues. Group performance degraded after denoising, collapsed under session-held-out evaluation, and remained partially recoverable from residual-noise representations when recording information leaked across random splits. These findings demonstrate how session-level confounding can produce misleadingly high classification performance in opportunistic bioacoustic datasets and highlight the importance of confound evaluation protocols in bioacoustic machine learning workflows.

Authors with CRIS profile

Involved external institutions

How to cite

APA:

Hauer, C., Nöth, E., Dobmeier, T., Arias Vergara, T., Maier, A., Vester, H., & Bergler, C. (2026). Recording-Specific Confounding in Killer Whale Vocalization Classification: A Cautionary Tale for Bioacoustic Machine Learning. In Proceedings of the 2026 International Symposium ELMAR (pp. 339-342). Zadar, HR.

MLA:

Hauer, Christopher, et al. "Recording-Specific Confounding in Killer Whale Vocalization Classification: A Cautionary Tale for Bioacoustic Machine Learning." Proceedings of the 2026 International Symposium ELMAR, Zadar 2026. 339-342.

BibTeX: Download