List view
- No due date•0/12 issues closed
- No due date•0/2 issues closed
# Problem Statement *The current landscape of bat acoustic monitoring tools for data visualization and machine learning is limited to close source options. The major existing commercial platforms such as SonoBat and Kaleidoscope offer a range of feature-rich workflows, but are not freely available or open source. This is a significant barrier to ongoing ecological research and slows the adoption of modern machine learning principles. The pace of bat research is therefore dependent on the access to proprietary analysis and results in the limited availability of public acoustic datasets.* # Proposed Contributions Replicating the success of other ecological endeavors in computer vision [1, 2], we propose to build an open source Python tool to analyze bat acoustic signals. The collection of acoustic samples for bat conservation consists mainly of the deployment of high-frequency microphones and recorders in remote and isolated locations, which record for extended periods of time between visits by field staff. The tool will focus on addressing the common and fundamental workflow constraint of finding relevant data (e.g., acoustic bat observations) in raw data collections. A major existing bottleneck with recording bat acoustic signatures is eliminating portions of the signal that contain nothing of interest (i.e., "dead air") or signals that are not ultrasonic emissions from bats. Similarly to the MegaDetector, which focuses on the task of eliminating non-relevant images from stationary camera trap installations, our tool will focus on eliminating non-relevant audio sequences from stationary acoustic capture stations. The proposed work will include the following contributions: - A public, open source Python implementation of high sample rate bat waveform loading and visualization with spectrograms - A published Python package on the PyPI (Python Package Index) that is easily installable on macOS, Windows, and Linux - A heuristic-based method for identifying the spectrogram frequencies/times that are too low energy to contain a bat pulse - An ONNX-optimized machine learning model for bat detection based on spectrograms, using CPU or any available GPUs - A terminal CLI and Python API for batch processing large collections of high sample rate waveform files - A collection of example Jupyter notebooks and documentation for researchers and developers to run local recordings - A submission procedure to allow researchers to contribute their acoustic data and annotations to the BatAI community At a high level, the tool will take in a folder of raw waveforms and save a new folder of filtered and relevant waveforms that contain only bat signals. The workflow will be simplistic and integrate with the BatAI platform for further analysis. Exposure of the core machine learning tools of the BatAI program to a wide audience of bat acoustic researchers will streamline local evaluation and increase trust in the platform's technology stack. The package will also be the basis for a published research paper on bat acoustic machine learning. [1] S. Beery, D. Morris, and S. Yang, "Efficient Pipeline for Camera Trap Image Review," arXiv:1907.06772, 2019. [2] "WildMeOrg/scoutbot," GitHub, Apr. 21, 2025. [Online]. Available: https://github.com/WildMeOrg/scoutbot
Overdue by 1 month(s)•Due by November 30, 2025•2/6 issues closed