BeeVe: Unsupervised Acoustic State Discovery in Honey Bee Buzzing
A novel unsupervised method named BeeVe, detailed in arXiv:2605.07903, identifies acoustic states in the buzzing of honey bees without the need for labels. This approach employs a frozen self-supervised Patchout Spectrogram Transformer (PaSST) to extract features, subsequently training a Vector-Quantized Variational Autoencoder (VQ-VAE) on these embeddings to create a discrete codebook of acoustic tokens from unannotated hive recordings. The process does not involve labels, pretext tasks, or contrastive objectives. A post-hoc assessment against established queen status reveals that the generated tokens effectively differentiate between queenright and queenless scenarios, yielding Jensen-Shannon Divergence values ranging from 0.609 to 0.688.
Key facts
- BeeVe is an unsupervised framework for acoustic state discovery in collective honey bee buzzing.
- It uses a frozen self-supervised Patchout Spectrogram Transformer (PaSST) as a feature extractor.
- A Vector-Quantized Variational Autoencoder (VQ-VAE) is trained without labels on PaSST embeddings.
- The framework learns a finite discrete codebook of acoustic tokens directly from unlabelled hive audio.
- No labels, pretext tasks, or contrastive objectives are used at any stage.
- Post-hoc evaluation against known queen status reveals token separation of queenright and queenless conditions.
- Jensen-Shannon Divergence values between 0.609 and 0.688 are achieved.
- The work addresses a gap in bioacoustic methods for non-vocal species.
Entities
—