Data, Supporting Information and Software for "Individual Bird Identification by Modeling Temporal Structure in Bioacoustic Embeddings"

Description This record contains the complete data and software package for the study "Individual Bird Identification by Modeling Temporal Structure in Bioacoustic Embeddings". It is intended to enable full reproducibility of the proposed workflow, covering data preprocessing, feature extraction, and model evaluation. This archive serves as the permanent snapshot for the code and data associated with the manuscript submitted to Methods in Ecology and Evolution. Contents Software & Source Code emb_extraction.zip: Snapshot of the preprocessing pipeline. Includes scripts for audio normalization, silence trimming, and embedding extraction using BirdNET backbone. emb-to-indiv.zip: Snapshot of the modeling pipeline. Includes the PyTorch implementations of the LSTM architecture and fully connected baselines, training loops, ablation study logic, and evaluation scripts. Documentation & Analysis Supporting Information S1: Detailed description of the additional trimming procedure applied to long-duration great spotted kiwi clips, including parameters and worked examples (input vs trimmed output). Supporting Information S2: Complete performance tables (macro-F1, accuracy, ROC-AUC) for all evaluated configurations, organized by species and experimental subset. Kiwi Vocalization Trimming (Jupyter notebook): Implementation used for development and visual quality control of the kiwi trimming pipeline. Data & Features Datasets/: Audio datasets repackaged into a standardized folder structure required to run the notebooks and scripts. All included audio data originate from sources released under Creative Commons Attribution 4.0 International (CC BY 4.0). Metadata/: Experimental metadata required to reproduce preprocessing and evaluation steps, including dataset split files and any corrected or consolidated metadata tables. Embeddings_from_3sPadding/: Precomputed BirdNET embeddings extracted from audio clips after padding durations to multiples of 3 seconds. Embeddings are stored as Parquet files (partitioned “parts”) to reduce computational overhead and enable direct reproduction of downstream experiments without re-running feature extraction. Embeddings_from_5sPadding/: Precomputed Perch embeddings extracted from audio clips after padding durations to multiples of 5 seconds. Embeddings are also stored as Parquet files (partitioned “parts”) to support baseline comparisons using the Perch feature representation. Licensing and attribution Audio Data: Provided under CC BY 4.0. Users must provide appropriate attribution to the original sources and indicate if any modifications were made. Where applicable, original dataset DOIs should be cited alongside this record. Precomputed embeddings distributed in this archive are derived from CC BY 4.0licensed audio material. As such, redistribution of embeddings requires appropriate attribution to the original audio sources. Software: The source code included in this archive is provided under the MIT License (see LICENSE files inside the respective zip archives). Reproducibility noteThe datasets are included in a project-ready layout to minimize setup overhead and avoid ambiguity in file organization. The metadata folder defines the experimental splits and any auxiliary information required to reproduce the reported results.