⚡ electrical engineering

ChiroEcho: extending automated bat vocalisation classification beyond the learned taxonomy

The paper introduces ChiroEcho, a deep learning framework that extends automated bat vocalization classification beyond its trained taxonomy by combining genus-level predictions with geographic distribution data to accurately identify species absent from the training set, thereby increasing operational coverage of native European bats from 73% to 85%.

Original authors: Burooj Ghani, Welmoed Eversteijn, Milan van Hirtum, Juan Sebastián Cañas, Vincent J. Kalkman, Dan Stowell, A. Leonie Baier

Published 2026-08-20

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Burooj Ghani, Welmoed Eversteijn, Milan van Hirtum, Juan Sebastián Cañas, Vincent J. Kalkman, Dan Stowell, A. Leonie Baier

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Bats are the silent guardians of the night, invisible to human eyes but vital to the health of ecosystems across Europe. They control insect populations, move nutrients, and help regulate the food web, yet their nocturnal habits make them incredibly difficult to study. Because they are also under strict legal protection and sensitive to environmental changes, scientists need reliable ways to track their populations over time. The most effective method for this is passive acoustic monitoring, where autonomous recorders listen for the high-pitched echolocation calls bats use to navigate and hunt. However, turning these recordings into useful data is a major challenge. Bat calls vary wildly depending on what the animal is doing, where it is flying, and even the weather, causing the sounds of different species to overlap and blur together. While computers have become good at identifying birds by their songs, teaching them to distinguish between similar-sounding bats has remained a stubborn problem, often limited to recognizing only the specific species the computer was explicitly taught to know.

A team of researchers has developed a new approach that breaks through this limitation by teaching a computer to listen not just for species, but for broader family groups, and then using a map to fill in the gaps. Their system, named ChiroEcho, is a deep learning framework trained on recordings of 35 different European bat species. Instead of trying to force the computer to memorize every single species immediately, the model first learns to identify the genus, or the broader family, to which a bat belongs. This is a crucial step because while two different species might sound nearly identical, they often belong to the same family. Once the computer makes a confident guess about the family, it checks the location where the recording was made. If the computer determines the bat belongs to a specific family, and the map shows that only one species of that family lives in that specific region, the system can confidently identify the species, even if it never heard that particular bat's voice during its training.

This method effectively allows the computer to recognize bats it has never encountered before. In their experiments, the researchers tested this by hiding two specific bat species from the training data entirely. When the system encountered recordings of these hidden species, the species-detection part of the model naturally guessed the wrong animal, usually picking a sound-alike relative that it had seen before. However, when the researchers applied the geographic rule, the system corrected itself. By combining the family-level guess with the known location, the system successfully identified the hidden species with high accuracy, recovering labels that were completely unavailable to the standard part of the model. This strategy expanded the system's ability to operate across Europe, increasing the number of native species it could identify from 35 to 41 out of the 48 native species in the region.

The researchers also discovered that relying solely on large datasets can sometimes hide the truth about how well a system works for rare animals. When they looked closely at the performance for species with very few recorded examples, the results were unstable and unpredictable. A single correct or incorrect guess could swing the success rate from zero to perfect, making it difficult to know if the computer was truly learning or just getting lucky. This highlights a persistent problem in conservation technology: the bats that need the most monitoring are often the ones with the least amount of data available to train the computers. The new framework does not solve this data shortage, but it offers a way to make the most of the data that does exist. By using the family-level prediction as a safety net and letting geography do the final work, the system can provide reliable identifications for a wider range of species without needing to be retrained for every new discovery.

The success of this project suggests a broader path forward for automated wildlife monitoring. It demonstrates that combining what a machine learns from sound with what humans know about where animals live can create a more powerful tool than either could be alone. The system does not need to be a perfect mimic of a human expert; instead, it acts as a partner that handles the heavy lifting of sound analysis while using simple, transparent rules to extend its knowledge. This approach could be applied to other animals, using different types of ecological knowledge, such as seasonal patterns or specific habitat associations, to help computers identify species they have never seen. For the bats of Europe, this means a future where automated monitoring can cover more ground and protect more species, turning the silent, dark world of the night into a landscape that is finally understood.

Technical Summary: ChiroEcho

Problem Statement
Automated monitoring of European bat populations is critical for biodiversity conservation due to their cryptic nocturnal lifestyles and legal protections. However, automated identification via passive acoustic monitoring remains challenging. Bat echolocation calls vary significantly based on behavior, habitat, and foraging context, leading to substantial overlap among species and within species. While deep learning has improved acoustic recognition, most existing systems are "closed-set," restricting predictions to a fixed taxonomy defined during training. This limits their applicability in real-world scenarios where novel or geographically restricted species may occur but lack sufficient training data. Furthermore, performance estimates for sparsely represented species are often unstable due to long-tailed data distributions in bioacoustic repositories.

Methodology
The authors present ChiroEcho, a deep learning framework designed to extend automated bat classification beyond its learned taxonomy by integrating ecological domain knowledge at the inference stage.

Dataset and Preprocessing: The model is trained on ChirosetEurope2, a curated dataset of 11,517 recordings covering 35 European bat species from 32 countries. The dataset exhibits a natural long-tailed distribution. Recordings are preprocessed into 20-second segments, resampled to 32 kHz, and converted to log-mel spectrograms (128 bins, 60 Hz–16 kHz). A noise class is constructed from non-bat environmental sounds to simulate field conditions.
Model Architecture: ChiroEcho utilizes a shared EfficientNet-B3 acoustic encoder with two parallel classification heads:
1. A Species Head predicting one of the 35 trained species.
2. An Auxiliary Genus Head predicting one of 11 genera.
  The model is trained end-to-end using a joint loss function ( $\mathcal{L} = \mathcal{L}_{\text{species}} + \mathcal{L}_{\text{genus}}$ ) with binary cross-entropy. Two initialization strategies are evaluated: standard ImageNet-1k weights (ChiroEchoIN) and weights from Perch 2.0 (a bioacoustic model trained on diverse taxa).
Geographic Resolution Strategy: The core innovation is an inference-stage rule that extends the operational taxonomy. When a recording is assigned to a specific genus with confidence above a threshold ( $\tau$ ), the system queries a geographic lookup table containing region–genus mappings. If the recording's location corresponds to a region where only one species of that predicted genus exists, the system resolves the prediction to that specific species, even if that species was absent from the training data. This transforms geographic priors from a constraint into a mechanism for taxonomy extension.
Evaluation Protocol: The authors evaluate closed-set performance (F1-score, mAP, AUROC) on the 35 trained species. They also conduct a controlled held-out experiment where two species (Pipistrellus kuhlii and P. maderensis) are removed from the training set and species-head label space. The system's ability to recover these "unseen" species via the geographic resolution rule is then tested.

Key Results

Closed-Set Performance: Both initialization strategies performed comparably. ChiroEchoIN (ImageNet-1k) achieved the highest mAP (0.694 species, 0.923 genus) and AUROC (0.990 species, 0.992 genus), while ChiroEchoPerch achieved slightly higher F1-scores (0.845 species, 0.886 genus). Genus-level metrics consistently outperformed species-level metrics, reflecting the difficulty of distinguishing closely related species.
Instability in Sparse Data: Analysis of species with fewer than 10 test segments revealed high instability in performance estimates. Average Precision (AP) varied significantly when combining validation and test splits (e.g., Myotis emarginatus AP shifted from 0.004 to 0.335), highlighting the difficulty of reliably evaluating rare species with limited data.
Geographic Resolution Success: In the held-out experiment, the species head naturally misclassified unseen species as acoustically similar congeners (e.g., P. kuhlii confused with P. nathusii). However, the geographic resolution rule successfully recovered the correct species for the held-out taxa:
- P. kuhlii (held out, tested in Pantelleria context): 69.5% recording-level accuracy.
- P. maderensis (held out, tested in Madeira context): 95.4% recording-level accuracy.
Taxonomic Coverage: By applying the geographic resolution rule to the 35 trained species, the framework's operational coverage extends from 35 to 41 of the 48 native European bat species, increasing coverage from 73% to 85%.

Significance and Claims
The paper claims to present the broadest operational coverage reported to date for automated European bat classification. Its primary significance lies in reframing geographic information not as a constraint that limits predictions to a fixed set, but as a mechanism to extend the effective taxonomy of a classifier.

The authors demonstrate a "resolve-when-unambiguous" strategy where coarse predictions (genus) combined with transparent external constraints (geographic distribution) can uniquely determine fine-grained classes (species) that were never seen during training. This approach offers a proof of principle for extending hierarchical classifiers without retraining or increasing architectural complexity. The authors note that this strategy is not limited to bats or geography; it could apply to any hierarchical classification problem where a reliable coarse prediction and an external domain constraint (e.g., seasonal occurrence, host association) uniquely identify a finer class.

The paper remains modest regarding limitations, acknowledging that performance for rare species is constrained by data scarcity and that the current framework relies on curated recordings, requiring future validation on uncurated, real-world soundscapes to assess robustness to domain shift.

Technical Summary: ChiroEcho

More like this