Hi everyone, We just published orcAI, a machine learning tool that detects and classifies a broad repertoire of killer whale acoustic signals, including vocalisations and incidental sounds (e.g. pulsed calls, whistles, breathing, tail slaps) and outputs time-stamped annotations from raw audio. The open access paper in Marine Mammal Science is now online at https://doi.org/10.1111/mms.70083.
orcAI is trained on a data set of acoustic signals of Icelandic killer whales. You can use this tool as is on your own audio recordings (WAV format), and it can also be trained on your own recordings if you have sufficient annotated data for training. Installation of the model on your computer is straigthforward by downloading from github https://github.com/ethz-tb/orcAI. orcAI includes an easy to use command-line interface. We would be very interested in your experience with our tool. If you encounter any issues, please reach out or open an issue on GitHub! Kind regards, Cherine (on behalf of all co-authors) Cherine Baumgartner PhD candidate Theoretical Biology Group ETH Zurich, Switzerland Abstract Acoustic monitoring is an essential tool for investigating animal communication and behavior when visual contact is limited, but the scalability of bioacoustic projects is often limited by time-intensive manual auditing of focal signals. To address this bottleneck, we introduce orcAI—a novel deep learning framework for the automated detection and classification of a broad acoustic repertoire of killer whales (Orcinus orca), including vocalizations (e.g., pulsed calls, whistles) and incidental sounds (e.g., breathing, tail slaps). orcAI combines a ResNet-based Convolutional Neural Network (ResNet-CNN) with Long Short-Term Memory (LSTM) layers to capture both spatial features and temporal context, enabling the model to classify signals and to accurately determine their temporal boundaries in spectrograms. Trained on a comprehensive dataset from herring-feeding killer whales off Iceland, the framework was designed to be adaptable to other populations upon training with equivalent data. Our final model achieves up to 98.2% accuracy on test data and is delivered as an open-source tool with an easy-to-use command-line interface. By providing a ready-to-use model that processes raw audio and outputs annotations, orcAI serves as a useful tool for advancing the study of killer whale vocal behavior and, more broadly, for understanding marine mammal communication and ecology.
_______________________________________________ MARMAM mailing list [email protected] https://lists.uvic.ca/mailman/listinfo/marmam
