Effects of interpulse-interval variation on deep-learning classification of bat vocalizations
This study investigates how inter-pulse interval (IPI) variability affects the performance of deep learning models in classifying bat echolocation calls. To this end, natural and normalized IPI datasets were constructed, and controlled experiments were conducted using EfficientNet-B0 and PaSST architectures to systematically quantify the differential sensitivities of convolutional neural networks and Transformers to temporal context. The findings reveal that natural IPIs contribute minimally to species classification; however, models trained on normalized data exhibit a significant accuracy degradation when evaluated on natural recordings. These results underscore the critical importance of distributional alignment between training and testing data, offering a novel perspective for evaluating the cross-condition generalization capabilities of bioacoustic models.