WABAD-Europe and ESC50 datasets formatted for machine learning

This dataset contains BirdNET embeddings, true labels, and acoustice indices values computed from the European recordings of the WABAD dataset V1 (A World Annotated Bird Acoustic Dataset for Passive Acoustic Monitoring). WABAD dataset corresponding authors: Cristian Pérez Granados ([e-mail]), Esther Sebastián-González ([e-mail]).Since the WABAD dataset is regularly updated, it is advisable to access the original files here for further research: https://zenodo.org/records/17293588. The WABAD dataset is composed of one-minute audio files (.wav) with corresponding Audacity and Raven Pro annotations at the species level, including start/end times and low/high frequency bounds. Embeddings and labels were also computed for the ESC-50 dataset, which contains environmental sounds: https://github.com/karolpiczak/ESC-50. The two datasets were formatted for machine learning as part of the following studies: Bernard, C., McEwen, B., Cretois, B., Glotin, H., Stowell, D., & Marxer, R. (2025). Data-driven Sampling Strategies for Fine-Tuning Bird Detection Models. bioRxiv. 2025-10.https://www.biorxiv.org/content/10.1101/2025.10.02.679964v1. The ‘results.zip’ file contains the intermediate computation results used in the GitHub repository associated with the article: https://github.com/mim-team/PAM_data_sampling. McEwen, B., Bernard, C., & Stowell, D. (2025). Stratified Active Learning for Spatiotemporal Generalisation in Bioacoustic Monitoring. BioRxiv, 2025-09.https://www.biorxiv.org/content/10.1101/2025.09.01.673472v2. Data processing steps: Dataset curation. Random split of the one-minute audio files into training (40%), validation (10%) and test (50%) sets. Segmentation of audio files into 3-seconds chunks. Computation of BirdNET predictions, uncertainty scores, and embeddings using https://github.com/birdnet-team/BirdNET-Analyzer. Computation of acoustic indices with Scikit-maad https://scikit-maad.github.io/. Storage of results in python .pkl files.本数据集包含从WABAD数据集V1（被动声学监测用全球标注鸟类声学数据集，英文全称为A World Annotated Bird Acoustic Dataset for Passive Acoustic Monitoring）的欧洲录音样本中计算得到的BirdNET嵌入特征、真实标签与声学指数数值。 WABAD数据集的通讯作者为Cristian Pérez Granados（邮箱：[e-mail]）与Esther Sebastián-González（邮箱：[e-mail]）。由于WABAD数据集会定期更新，如需开展后续研究，建议通过以下链接获取原始文件：https://zenodo.org/records/17293588。 WABAD数据集由单分钟时长的音频文件（.wav格式）构成，配套有Audacity与Raven Pro软件生成的物种级标注信息，包含音频片段的起止时间以及低频、高频阈值范围。 本数据集还为ESC-50数据集计算了嵌入特征与标签，该数据集包含环境音效样本，相关开源链接为：https://github.com/karolpiczak/ESC-50。 上述两个数据集已针对机器学习任务完成格式适配，相关支撑研究如下： 1. Bernard, C., McEwen, B., Cretois, B., Glotin, H., Stowell, D., & Marxer, R. (2025). 面向鸟类检测模型微调的数据驱动采样策略. bioRxiv, 2025-10. https://www.biorxiv.org/content/10.1101/2025.10.02.679964v1. "results.zip"文件包含该论文关联的GitHub仓库（https://github.com/mim-team/PAM_data_sampling）中使用的中间计算结果。 2. McEwen, B., Bernard, C., & Stowell, D. (2025). 面向生物声学监测时空泛化的分层主动学习方法. BioRxiv, 2025-09. https://www.biorxiv.org/content/10.1101/2025.09.01.673472v2. 数据处理步骤如下： 1. 数据集整理：将单分钟音频文件随机划分为训练集（40%）、验证集（10%）与测试集（50%）。 2. 音频分割：将音频文件切割为3秒时长的片段。 3. 特征计算：通过https://github.com/birdnet-team/BirdNET-Analyzer计算BirdNET预测结果、不确定性得分与嵌入特征。 4. 声学指数计算：使用Scikit-maad工具包（https://scikit-maad.github.io/）计算声学指数。 5. 结果存储：将计算结果存储为Python .pkl格式文件。