PSTSE: Perch-SED Taxonomic Sequence Ensemble Multi-Taxa Acoustic Recognition in the Pantanal for BirdCLEF+ 2026

We developed PSTSE, a Perch-SED Taxonomic Sequence Ensemble, for multi-label recognition of 234 animaltaxa in five-second windows from the Pantanal. Our objective was practical: combine strong frozen bioacousticrepresentations with models that could use within-file temporal context and still complete offline CPUinference within Kaggle’s 90-minute limit. PSTSE uses Perch-v2 embeddings and logits, lightweight prototypeand state-space sequence heads, a distilled convolutional sound-event detector, and weak genus/class regularisation.Our private-best submission achieved public/private macro ROC-AUC values of 0.95035/0.94216,and we finished 947th of 4,094 teams. This paper explains the system we executed, the reasoning behindits architecture, its relationship to approaches reported by leading participants, and the limits of the evidencepreserved during the competition. Some backbones and starting implementations were public competition resources;our contribution lies in selecting, adapting and integrating them into the submitted pipeline. Becausewe did not preserve a controlled ablation for every operation, we report leaderboard differences as observationsrather than causal effects.