Research library

BioME: A Resource-Efficient Bioacoustic Foundational Model for IoT Applications

BioME is a family of resource-efficient bioacoustic audio encoders trained by layer-wise knowledge distillation and conditioned with modulation-spectrum features. The models reduce parameter count by up to 75 percent while matching or exceeding larger encoders on broad bioacoustic benchmarks. The authors also evaluate queen-presence classification, colony-strength regression, bee identification, and voice-activity detection from beehive audio. The six-million-parameter Edge model produced the best efficiency-weighted aggregate score and generalized strongly across hives. Code and pretrained checkpoints are public.

Download PDF

Publication details

Authors
Heitor R. Guimarães, Abhishek Tiwari, Mahsa Abdollahi, Anderson R. Avila, Tiago H. Falk
Organizations
🇨🇦 Institut national de la recherche scientifique (INRS)
Year
2026
Type
Preprint

Models and access

Relevancy to Gratheon

BioME is directly actionable for Gratheon's acoustic monitoring stack because it benchmarks one reusable encoder across four hive tasks rather than optimizing a separate feature pipeline for each outcome. The cross-hive queen-presence test is particularly important for avoiding models that memorize the acoustics of individual colonies or recording setups.

The six-million-parameter Edge checkpoint offers a concrete starting point for local inference and transfer learning. Its strong efficiency-weighted score and publicly downloadable model variants let Gratheon compare accuracy, memory, and latency trade-offs before collecting enough labeled hive audio to train a foundation model from scratch.