MyBAD: Malaysian Bird Activity Detection Dataset

MyBAD (Malaysian Bird Activity Detection) is a large-scale audio dataset for Edge AI bird detection, comprising 28,000 positive (bird-present) and 28,000 negative (bird-absent) 3-second clips sampled at 16 kHz. The dataset represents all available bird sound samples from Xeno-canto recordings in Malaysia and Singapore up to June 2025. Each clip is provided as a .npy (NumPy array) file derived from mel-spectrograms, with five different resolutions: 80×184, 64×184, 48×184, 32×184, and 16×184. The dataset is designed for automated bird activity detection, biodiversity monitoring, and real-time acoustic sensing on low-power devices, and is fully machine-learning–ready. Positive samples were primarily extracted from Xeno-canto recordings of Malaysian bird species by selecting the highest-energy 3-second segment from each source recording. Additional samples were drawn from the Macaulay Library to improve representation of common but underrecorded species.Negative samples originate mainly from the BirdCLEF “BAD” competition (BirdVox, Freefield1010, Warblr) and are supplemented with non-bird environmental audio from Xeno-canto, ESC, and FSC datasets. All negative clips were screened to ensure the absence of bird vocalizations. A future release will include full metadata (source IDs, duration, and license information) to further strengthen the traceability of each sample. Released under CC BY 4.0, MyBAD provides an open, reproducible resource for developing Edge AI models for tropical bird monitoring and acoustic biodiversity assessment. Note: This dataset will be formally described in a forthcoming paper introducing the dataset and companion featherweight bird-activity detector.