SEABAD: Southeast Asian Bird Activity Detection

SEABAD (Southeast Asian Bird Activity Detection) is a large-scale, curated audio dataset for tropical bird activity detection and passive acoustic monitoring (PAM). It contains 50,000 three-second WAV clips (16-bit PCM, 16 kHz mono), evenly split between 25,000 bird-present (positive) and 25,000 bird-absent (negative) samples, and is designed for edge AI inference on low-power embedded devices.                                                  

Dataset Composition

Positive samples span 1,677 Southeast Asian bird species sourced from Xeno-Canto recordings (through December 2025) across Malaysia (43.1%), Thailand (27.4%), Indonesia (22.0%), Singapore (7.2%), and Brunei (0.2%). A six-stage curation pipeline — metadata acquisition, download and resampling, acoustic deduplication, segment extraction, diversity-aware species balancing, and quality assurance — reduced class imbalance by 13.7% (Gini coefficient: 0.601 → 0.519). 92.1% of clips carry Xeno-Canto quality ratings of A or B. A manual audit of 1,000 randomly sampled clips confirmed 97.8% ± 0.9% labeling accuracy.

Negative samples (25,000 clips) are drawn from six open-access environmental audio datasets: BirdVox-DCASE-20k (9,983 clips), Freefield1010 (5,755 clips), Warblrb10k (1,950 clips), FSC-22 (1,875 clips), ESC-50 (1,840 clips), and DataSEC (3,597 clips). All avian classes were removed from each source before inclusion.

Baseline performance

MobileNetV3-Small (1.1M parameters), the primary edge-deployment baseline, achieves 99.57% ± 0.25% accuracy and AUC 0.9985 ± 0.0002 on the held-out test set (averaged across three random seeds). Three larger architectures (VGG16, ResNet50, EfficientNetB0) all exceed 99.4% accuracy, confirming high label quality and task separability.

Dataset splits                                                                                                                                                                                        

Training: 40,000 clips (80%) · Validation: 5,000 clips (10%) · Test: 5,000 clips (10%), stratified 50/50 positive/negative.

File note                                                                                                                                                                                                 

The archive is named mybad.zip for historical reasons (MyBAD was the working title during curation). The contents are the SEABAD dataset as described in the accompanying paper.                                                

License

All positive clips are sourced from Xeno-Canto under Creative Commons licenses (CC BY-SA, CC BY-NC-SA, CC BY-NC-ND, CC0). The dataset compilation is released under CC BY 4.0. Users must adhere to the respective source licenses when redistributing.

Related resources                                                                                                                                                                               

• Paper: SEABAD: A Tropical Bird Activity Detection Dataset for Passive Acoustic Monitoring — Zabidi, Idris MY, Idris N — https://arxiv.org/abs/2605.20853
• Curation code: https://github.com/mun3im/seabad