[Paper Review] BioME: A Resource-Efficient Bioacoustic Foundational Model for IoT Applications

Heitor R. Guimarães, Abhishek Tiwari+3 more|arXiv (Cornell University)|Feb 10, 2026

Animal Vocal Communication and Behavior0 citations

TL;DR

BioME is a resource-efficient self-supervised audio encoder for bioacoustic tasks, distilled from a large teacher model and enhanced with modulation-based features via FiLM to enable strong performance on edge devices and across diverse ecological tasks.

ABSTRACT

Passive acoustic monitoring has become a key strategy in biodiversity assessment, conservation, and behavioral ecology, especially as Internet-of-Things (IoT) devices enable continuous in situ audio collection at scale. While recent self-supervised learning (SSL)-based audio encoders, such as BEATs and AVES, have shown strong performance in bioacoustic tasks, their computational cost and limited robustness to unseen environments hinder deployment on resource-constrained platforms. In this work, we introduce BioME, a resource-efficient audio encoder designed for bioacoustic applications. BioME is trained via layer-to-layer distillation from a high-capacity teacher model, enabling strong representational transfer while reducing the parameter count by 75%. To further improve ecological generalization, the model is pretrained on multi-domain data spanning speech, environmental sounds, and animal vocalizations. A key contribution is the integration of modulation-aware acoustic features via FiLM conditioning, injecting a DSP-inspired inductive bias that enhances feature disentanglement in low-capacity regimes. Across multiple bioacoustic tasks, BioME matches or surpasses the performance of larger models, including its teacher, while being suitable for resource-constrained IoT deployments. For reproducibility, code and pretrained checkpoints are publicly available.

Motivation & Objective

• Motivate efficient bioacoustic representation learning for resource-constrained IoT deployments.
• Develop a small yet powerful encoder by distilling a large teacher model while preserving cross-domain generalization.
• Enhance representations with modulation-aware features to improve disentanglement in low-capacity regimes.
• Demonstrate BioME's effectiveness across BEANS bioacoustic benchmarks and acoustic beehive monitoring tasks.

Proposed method

• Patch-based audio encoding using mel-spectrograms and non-overlapping 16x16 patches.
• Transformer-based encoder with grouped query attention (GQA) and Rotary Position Embedding (RoPE) to reduce memory and compute.
• FiLM conditioning injects modulation-spectrum based context (MSAB features) at each layer to bias representations toward bioacoustic signals.
• Layer-wise knowledge distillation from BEATs as teacher to a thinner student with matching depth (12 layers) and intermediate alignment at layers {3,6,9,12} using an L1 plus cosine similarity loss.
• Multi-domain pretraining (speech, environmental sounds, bioacoustic recordings) to improve ecological generalization.

Figure 1: Spectrogram (top row) and the modulation spectrum (bottom row) plots, averaged across samples, for three hummingbird species: Anna’s hummingbird ( annhum ), Broad-tailed hummingbird ( brthum ), and Costa’s hummingbird ( coshum ).

Experimental results

Research questions

• RQ1 Can a compact student encoder match or surpass its large teacher on bioacoustic tasks through layer-wise distillation?
• RQ2 Does incorporating modulation-based features via FiLM improve representation disentanglement in low-capacity models?
• RQ3 How do architectural choices (GQA, RoPE, Llama-inspired components) affect efficiency and accuracy on edge devices?
• RQ4 What is BioME’s performance trade-off across BEANS benchmarks and beehive monitoring tasks when scaled for edge vs. base deployments?

Key findings

• BioME 6M (Edge) achieves state-of-the-art efficiency, outperforming baselines with up to 15x fewer parameters in BEANS tasks.
• BioME Small (26M) and Base (76M) surpass the teacher BEATs on BEANS overall scores, illustrating effective distillation and inductive biases.
• Modulation features (MSAB) via FiLM improve feature disentanglement, enabling the student to sometimes exceed the teacher on select tasks.
• Optimized spectral resolution (NFFT=256) balances classification and detection, achieving strong performance across tasks.
• BioME demonstrates strong edge performance for acoustic beehive monitoring, with an inverse scaling trend where the smallest model can outperform larger variants on BSTS.

Figure 2: Block diagram of a single Transformer layer in the proposed BioME encoder. Patch embeddings are processed alongside side-channel context features, which are integrated at each layer through the Conditioner module implementing FiLM-based conditioning.