11th Workshop on Detection and Classification of Acoustic Scenes and Events

DCASE 2026 will take place in Boston from October 28 to 29, 2026. Fraunhofer IDMT will present its latest research findings in the field of sound event detection.

Abstract—In this study, we address the automated detection of whale vocalizations in low-sampling-rate passive acoustic monitoring (PAM) recordings from the BioDCASE Challenge 2026 Task 2. We investigate two spectrogram-based object detectors from the computer vision domain, YOLOv11 and RT-DETR, alongside an audio-native transformer model, BEATs, fine-tuned on upsampled waveforms. Using an event-level evaluation based on 1D intersection-over-union (IoU) along the temporal axis, we observe that the BEATs model delivers the highest performance among the base models on the test set (F1 = 0.655), indicating that large-scale audio pre-training transfers effectively even to  narrowband, low-frequency whale vocalizations. RT-DETR clearly surpasses YOLOv11 in its base configuration in the test set of the development set (F1 = 0.517 vs. 0.430), and a detailed analysis of model-agnostic post-processing reveals that adding a simple non-maximum suppression (NMS) step is essential to improve instance-level performance, increasing RT-DETR’s test-set F1 score to 0.679. On the evaluation set, the same overall ranking is observed, with BEATs achieving an F1 score of 0.376, RT-DETR with NMS reaching 0.276, and YOLOv11 with augmentation obtaining 0.166. These results demonstrate that (i) self-supervised transformer models pre-trained on large-scale audio datasets already provide strong baselines for marine bioacoustic event detection, and (ii) transformer-based detectors with carefully tuned NMS remain competitive for precise temporal localization, offering a complementary approach when frame-level spectrograms and high-resolution instance-level annotations are available.