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

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.

💡 Research Summary

The paper introduces BioME, a resource‑efficient self‑supervised audio encoder tailored for bioacoustic tasks on Internet‑of‑Things (IoT) edge devices. While recent SSL models such as BEATs and AVES achieve strong performance, their ~90 M parameter size and high inference cost make them unsuitable for low‑power, real‑time deployments. BioME addresses this gap through two complementary strategies.

First, the authors employ layer‑to‑layer knowledge distillation from a high‑capacity teacher (BEATs). The student mirrors the teacher’s depth (12 Transformer layers) but reduces hidden dimensions, achieving a 75 % reduction in total parameters. Distillation is performed at four intermediate layers (3, 6, 9, 12), with linear projection heads aligning student and teacher representations when dimensions differ. This approach enables the compact model to inherit the teacher’s rich feature space without sacrificing accuracy.

Second, BioME injects modulation‑spectrum information—a DSP‑derived representation that captures the rate of change of spectral components—into every Transformer layer via Feature‑wise Linear Modulation (FiLM). The modulation spectrogram is summarized into Modulation Spectrogram Average Bands (MSAB), a 512‑dimensional vector obtained by averaging across modulation‑frequency and acoustic‑frequency axes. FiLM computes per‑channel scale (γ) and shift (β) parameters from the MSAB vector and applies them to the patch embeddings (x′ = γ ⊙ x + β). This lightweight conditioning provides a top‑down inductive bias that helps the model disentangle animal vocalizations from environmental noise, especially in low‑capacity regimes.

The front‑end follows a BEATs‑style pipeline: raw waveforms are converted to mel‑spectrograms, split into non‑overlapping 16 × 16 patches, and linearly projected. The Transformer backbone incorporates several efficiency‑focused components: Grouped Query Attention (GQA) reduces key‑value projection parameters, Rotary Position Embedding (RoPE) eliminates separate positional embeddings, SiLU activation, and RMSNorm normalization. These choices further lower memory usage and improve training stability on constrained hardware.

Training data comprises a multi‑domain corpus of 5,200 hours, including BioAudioSet (4,620 h), FSD50K, VGGSound, and iNatSounds. This diverse pre‑training set is intended to promote cross‑domain generalization, exposing the model to speech, environmental sounds, and a wide range of animal vocalizations. After self‑supervised pre‑training, BioME is fine‑tuned on the BEANS benchmark (10 bioacoustic classification/detection tasks covering birds, whales, bats, insects, and mammals) and on a beehive‑monitoring suite (queen presence, colony strength, buzzing classification, activity detection).

Experimental results show that BioME matches or slightly exceeds the teacher’s performance on average accuracy and F1‑score across the benchmark while using roughly one‑quarter of the parameters and significantly fewer FLOPs, enabling near‑real‑time inference on typical IoT hardware. Ablation studies confirm the importance of each component: removing FiLM or the MSAB features degrades performance by 2–3 %, and omitting GQA or RoPE increases memory consumption without notable accuracy gains, highlighting their role in efficiency.

The authors acknowledge limitations: computing the modulation spectrogram adds preprocessing overhead; the current design is fixed to 12 layers, leaving the effectiveness of deeper or shallower student architectures unexplored; and ultra‑low‑power microcontrollers may still face memory constraints. Future work aims to learn modulation features end‑to‑end with a lightweight front‑end, explore asymmetric teacher‑student configurations, and further compress the model for deployment on sub‑gram‑scale devices.

In summary, BioME demonstrates that a carefully distilled, modulation‑aware encoder can deliver state‑of‑the‑art bioacoustic performance within the strict computational budget of IoT edge nodes, paving the way for scalable, continuous wildlife monitoring.