An Open Reproducible Framework for CNN-Based Cetacean Vocalization Detection in Passive Acoustic Monitoring
Convolutional neural networks (CNNs) operating on spectrogram images are the established method for automated cetacean whistle detection in passive acoustic monitoring (PAM). The FFT window length (N_fft) used during spectrogram generation is routinely fixed without justification, yet it determines the time-frequency resolution and, consequently, how the resulting image is distorted when resized to a fixed CNN input dimension. This study presents a controlled sensitivity analysis of N_fft across five values (128, 256, 512, 1024, 2048) on binary Tursiops truncatus whistle detection, using 10-fold cross-validation on an in-domain dataset (Oltremare, 192 kHz) and cross-domain evaluation on an independent open-ocean benchmark (DCLDE 2022). All experiments were conducted within a formally defined, open-source pipeline (ai-pam-pipeline). In-domain performance is uniformly high across all configurations. Cross-domain results diverge: N_fft = 256 significantly outperforms 512, 1024, and 2048 in macro F1, while maintaining a false discovery rate (FDR) of exactly zero across all 10 folds and all tested classification thresholds. N_fft = 128 achieves comparable recall but produces FDR > 0 in all 10 folds, exhibiting a transfer failure that is undetectable by in-domain validation. These results show that a preprocessing parameter routinely treated as an implementation detail has a large, systematic effect on cross-domain generalization. The study does not aim to propose N_fft = 256 as a universal optimum; rather, it tests the common assumption that this parameter is neutral and safe under a fully specified representation regime, and demonstrates that it is not.