An open reproducible framework for CNN-based cetacean vocalization detection in passive acoustic monitoring

Background Passive acoustic monitoring (PAM) combined with deep learning is increasingly used to detect cetacean vocalizations. The workflow from raw audio to a trained detector — annotation, segmentation, spectrogram parameterization, image processing, and model evaluation — is seldom accompanied by the complete set of configuration parameters, code, and data needed for independent reproduction, limiting methodological comparability and systematic benchmarking. Methods This article presents ai-pam-pipeline, an open-source six-stage framework that formalizes the complete CNN-based detection workflow, with every parameter defined in a single YAML configuration file ensuring exact experimental reproducibility. The framework separates methodological structure from computational implementation, allowing single-channel, spectrogram-based classification workflows to be adapted to other species and vocalization types through configuration alone, without code modification. Validation was conducted on bottlenose dolphin ( Tursiops truncatus ) recordings through two complementary experiments: (A) a binary whistle detector trained on controlled-environment data and evaluated on an independent open-ocean dataset, and (B) a five-class vocalization classifier demonstrating framework scalability. Results The binary detector achieved macro F1 = 0.960 under ten-fold cross-validation with file-level grouped splitting. Cross-domain evaluation yielded a precision of 0.999 and a false discovery rate of 0.001 (macro F1 = 0.744); precision-recall analysis shows that recall rises to 0.76 at a lower decision threshold while precision remains above 0.95, indicating partial but operationally useful generalization. The multiclass extension achieved macro F1 = 0.835 across five vocalization categories; inter-class confusion between echolocation click trains and burst-pulse sounds reflects biological signal overlap rather than classifier failure. Conclusions The framework provides a reproducible and documented baseline for CNN-based cetacean detection in PAM. The complete toolkit is released under the Apache 2.0 license, with training and evaluation datasets made openly available through separate public repositories. Together, these resources enable exact replication of the reported analyses and support adaptation to new monitoring contexts.