Building High-quality Corpus for Sound of Wild Animals

With rapid advances in acoustic sensing technologies and data analysis algorithms, acoustic monitoring is transitioning from traditional ecological surveys to an AI-driven stage. Despite the exponential growth of animal acoustic data, high-quality, uniformly annotated and verified datasets compatible with AI model training remain scarce.

Here we present a comprehensive methodology for constructing a wildlife sound dataset tailored to AI model development, covering multi-source heterogeneous data collection, standardised preprocessing pipelines, a quality control system, and a quantitative evaluation framework. The current release covers three taxonomic groups: Aves, Amphibia, and Insecta, comprising 44 species, 108 audio recordings with matching annotation files, 1,036 annotated bounding boxes, and 1,036 spectrogram images.