BeeVe: Unsupervised Acoustic State Discovery in Honey Bee Buzzing

Researchers introduce BeeVe, an unsupervised machine learning framework that discovers acoustic patterns in honey bee hive sounds without labels or predefined categories. The system successfully identifies distinct behavioral states linked to hive health conditions, demonstrating that AI can extract meaningful biological structure from non-vocal animal signals.

BeeVe represents a significant advance in computational biology by tackling the challenge of extracting actionable information from unlabeled biological signals. Traditional bioacoustic analysis relies on species-specific vocal production models or human-annotated categories, approaches that fail for collective organisms like bee colonies. This work circumvents those limitations by combining frozen feature extraction via the Patchout Spectrogram Transformer with unsupervised discrete token learning through VQ-VAE, eliminating the need for labeled training data entirely.

The framework's ability to distinguish queenright from queenless conditions—states with significant implications for colony survival and productivity—validates that the learned acoustic tokens capture behaviorally relevant information. The discovery that queenless conditions decompose into three internally coherent sub-states suggests the system captures granular shifts in hive dynamics that would require domain expertise to annotate manually. Token transition analysis revealing non-random sequential structure indicates that the codebook captures not just static sounds but meaningful behavioral sequences.

For the agricultural and apiculture sectors, this approach enables non-invasive, scalable hive monitoring. Current hive health assessment requires manual inspection and expertise; acoustic monitoring could democratize early detection of colony dysfunction, disease, or queen loss. The strong generalization metrics—including 0.947 Jaccard overlap on unseen recordings—suggest practical deployment feasibility. This work also establishes a methodological template for studying other collective biological systems lacking clear vocalization models, from insect colonies to marine organisms, expanding AI's reach into domains where supervised learning has been impractical.

• →BeeVe uses self-supervised learning to discover discrete acoustic states in bee hives without any manual labels or predefined categories.
• →The system successfully identifies queen presence/absence conditions and decomposes queenless states into three internally coherent sub-states.
• →Learned acoustic tokens show strong generalization to unseen recordings with 94.7% Jaccard overlap, indicating practical applicability.
• →Non-random token transitions (p << 0.001) confirm the framework captures behaviorally meaningful sequential structure, not random patterns.
• →This approach enables scalable, non-invasive hive health monitoring and establishes methods for studying other collective biological systems.