SEA-BIRD30: a machine learning–ready dataset for 30 Southeast Asian bird sounds

SEA-BIRD30 is a curated, machine-learning-ready collection of avian vocalizations from 30 bird species found across Southeast Asia, released as a follow-up to the original ten-species SEA-Bird dataset. Recordings are sourced primarily from the Southeast Asian region — longitude 60-140°E, or residence in an ASEAN member country — spanning a broader range of habitats than its predecessor, from garden and urban species to forest and nocturnal birds such as Large-tailed Nightjar, Greater Racket-tailed Drongo, and Pygmy Cupwing. The dataset comprises 18,000 three-second audio clips sampled at 16 kHz (600 clips per species), each manually verified to ensure minimal background noise, balanced class representation, and the absence of overlapping species calls. SEA-BIRD30 extends the original SEA-Bird dataset's goal of addressing the scarcity of machine-learning-ready acoustic datasets from underrepresented tropical regions, tripling species coverage while maintaining the same curation standard. It supports research in automated species identification, biodiversity monitoring, and embedded acoustic sensing, with its 16 kHz sample rate suited to Edge AI applications on low-power microcontrollers. All recordings were sourced from the Xeno-canto open repository under Creative Commons licenses and subjected to a rigorous curation process, including segmentation and manual inspection of spectrograms to ensure vocal clarity and completeness. The dataset is organized as one folder per species (e.g., Common_Tailorbird/, Spotted_Dove/), each containing that species' clips. Individual clips are named using the original Xeno-canto recording ID followed by the clip start time in milliseconds relative to the source recording (e.g., xc123456_15000.wav). This naming and directory structure preserves traceability and facilitates seamless integration into automated processing pipelines. A companion train/validation/test split is provided as a CSV file mapping each clip to its assigned split, rather than via separate split directories. The training, validation, and test splits follow an exact 80/10/10 ratio (14,400 / 1,800 / 1,800 clips) and were generated using mixed-integer programming to eliminate data leakage — no source recording contributes clips to more than one split. Five lightweight-to-mid-sized convolutional architectures were trained and evaluated on this exact 30-species split to validate the dataset's suitability for deep-learning applications: EfficientNet-B0 reached 75.61% and MobileNetV3-Small reached 71.67%. On custom audio-oriented MynaNet-1o (64-mel front end, 122.8K parameters) reached 89.39% FP32 / 89.50% INT8 test accuracy; MynaNet-1q (64-mel) reached 89.13% / 89.02%; MynaNet-1q (16-mel) reached 85.39% / 85.56%; TailorNet (37.7K parameters) reached 84.87%. These results confirm the dataset supports strong closed-set classification performance even from sub-50K-parameter models suitable for microcontroller deployment. See https://github.com/mun3im/mynanet SEA-BIRD30 is released under a Creative Commons Attribution (CC BY 4.0) license and is freely available for academic and applied research. The dataset contributes to the growing body of open bioacoustic resources that promote reproducible, scalable, and inclusive research in tropical biodiversity informatics.