ChiroEcho: extending automated bat vocalisation classification beyond the learned taxonomy
Authors: Burooj Ghani, Welmoed Eversteijn, Milan van Hirtum, Juan Sebastián Cañas, Vincent J. Kalkman, Dan Stowell, A. Leonie Baier
Organizations: Naturalis Biodiversity Center · Department of Cognitive Science and Artificial Intelligence, Tilburg University · People and Nature Lab, University College London · Leiden Institute of Advanced Computer Science, Leiden University
Bats are key indicators of ecosystem health and are protected throughout Europe, making reliable population monitoring a conservation priority. Their cryptic nocturnal lifestyle makes passive acoustic monitoring essential, yet automated identification remains difficult as echolocation calls vary with behaviour and environment and overlap among species. We present a deep learning framework that jointly predicts species and genus and combines genus predictions with geographic species distributions at inference. When only one species of a predicted genus occurs in a region, the framework can resolve species absent from the learned taxonomy. This reframes geographic information as a means of extending, rather than constraining, a classifier's effective taxonomy. Using recordings spanning 35 European bat species, we evaluate closed-set classification, examine the instability of performance estimates for sparsely represented species, and conduct a controlled held-out proof-of-principle experiment. The rare-species analysis shows how limited evaluation data can obscure species-level performance, while the held-out experiment shows that genus predictions and location can recover labels unavailable to the species head. Geographic resolution extends operational coverage from 35 to 41 of the 48 native European bat species, increasing coverage from 73% to 85%. To our knowledge, this is the broadest operational coverage reported for automated European bat classification. More broadly, the bat framework provides proof of principle for resolving unseen fine-grained classes by combining coarse predictions with transparent external constraints.
Figures & tables
Figure 1: Illustration of inference-stage geographic resolution using the example of Plecotus kolombatovici . (a, c) Echolocation call sequences of P. austriacus and P. kolombatovici , respectively. (b) Approximate Southern European distribution ranges of P. austriacus (orange) and P. kolombatovici (blue), with overlap shown by orange–blue hatching (ranges based on Dietz et al. [ 13 ] ). Owing to insufficient acoustic recordings, P. kolombatovici is not included in the learned taxonomy. The species head may therefore assign such a recording to an acoustically similar congener class, illustrated here by P. austriacus . For a recording from Cyprus, a genus-level Plecotus prediction above the confidence threshold activates the geographic lookup ( Table 1 ), which resolves the recording to P. kolombatovici because it is the sole regional representative of the genus.
Figure 2: Schematic overview of the ChiroEcho architecture and geographic-resolution step, illustrated for Myotis punicus in Malta. The depicted species-head output, M. blythii , is illustrative rather than observed and represents a plausible confusion with a closely related species whose echolocation characteristics overlap.
Genus
Region
Resolved species
In-region data
Taxonomy extension
Cnephaeus
Gibraltar
C. isabellinus
✓
Myotis
Malta
M. punicus
✓
Myotis
Pantelleria
M. punicus
✓
Nyctalus
Azores
N. azoreum
✓
Plecotus
Malta
Pl. gaisleri
✓
Plecotus
Pantelleria
Pl. gaisleri
✓
Table 1: Region–genus lookup table used to resolve species s from predicted genus g^ and location ℓ ( Algorithm 1 ). A genus prediction is resolved only when a single species of that genus occurs in the listed region. “In-region data” indicates that labeled recordings of the resolved species were available from that region. “Taxonomy extension” indicates that the resolved species was absent from the training taxonomy.
Model
Weights initialization
Species
Genus
F1-score
mAP
AUROC
F1-score
mAP
AUROC
ChiroEcho IN
ImageNet-1k
0.836
0.694
0.990
0.880
0.923
0.992
ChiroEcho Perch
Perch 2.0
0.845
0.672
0.988
0.886
0.921
0.988
Table 2: Recording-level closed-set species and genus classification performance of ChiroEcho, trained on 35 European bat species and one noise class with ImageNet-1k or Perch 2.0 initialization. Segment-level predictions are aggregated to recording-level predictions by max pooling before evaluation. Best performance per metric is shown in bold.
Species
# Segments
Average precision
Test
Val+Test
Test
Val+Test
Myotis davidii
1
2
1.000
1.000
Myotis alcathoe
1
11
0.019
0.013
Myotis blythii
2
15
0.001
0.001
Myotis mystacinus
2
42
0.019
0.074
Myotis emarginatus
4
7
0.004
0.335
Table 3: Classification results for species represented by fewer than 10 test-split segments using ChiroEcho IN . Average precision (AP) is reported for both the test split and the combined validation+test split.
Species
# Segments
Species head (baseline)
Geo-resolved species accuracy
Top confusion
%
P. kuhlii
803
P. nathusii
37.4%
0.695
P. maderensis
44
P. pipistrellus
93.2%
0.954
Table 4: Geographic resolution of two Pipistrellus species held out entirely from the training taxonomy and species-head label space. All available segments were used for evaluation. Because neither species has an output class, direct identification by the species head is impossible; the “Top confusion” columns instead report the most frequent segment-level top-1 prediction among the trained classes and its share of segments. Geo-resolved species accuracy is reported at the recording level after max pooling and application of the geographic resolution rule ( Algorithm 1 ).
Scientific name
CSE
No.
Distribution within EUROBATS area
Barbastella barbastellus
✓
45
Widespread across Europe and adjacent western Asia
Barbastella caspica ∗
4
Caucasus and Caspian region
Cnephaeus anatolicus ∗
5
Eastern Mediterranean
Cnephaeus isabellinus
6
Iberia and North Africa
Cnephaeus nilssonii
✓
30
Northern, central and eastern Europe
Cnephaeus ognevi
6
Caucasus and Middle East
Table S1: The 55 bat taxa listed under the EUROBATS Agreement on the Conservation of Populations of European Bats [ 56 ] , presented using current nomenclature, and their representation in ChirosetEurope (‘CSE‘). ‘No.‘ indicates the number of EUROBATS Range States ( n=60 ) in which a species is listed as present. Species marked with ∗ have no documented occurrence within the territory covered by the latest EU Action Plan for Bats (2018–2024) [ 3 ] . Data current as of July 2026.
Learned species
Train
Val
Test
Total
Duration
Countries
Barbastella barbastellus
170
77
37
284
00:46:29
14
Cnephaeus nilssonii
79
54
44
177
00:34:44
12
Cnephaeus serotinus
253
123
79
455
01:51:08
17
Hypsugo savii
399
100
160
659
01:00:08
14
Miniopterus schreibersii
112
109
30
251
00:40:56
6
Myotis alcathoe
5
2
1
8
00:00:45
5
Table S2: Per-species composition and operational taxonomic coverage of ChiroEcho . For the 35 species represented in ChirosetEurope, the table reports recording counts by split, total audio duration, and the number of recording countries. The lower block lists six additional species supported in ChiroEcho through geographic resolution and the regions in which genus and location uniquely determine species identity. Dataset statistics are not applicable to these species because they are outside the learned taxonomy.
Figure S1: Recording abundance and relative geographic coverage across the 35 species in ChirosetEurope. Species are ordered by decreasing recording count, represented by bar height. Bar colour indicates the proportion of each species’ reported EUROBATS range states represented by at least one recording, with darker shades indicating greater coverage. Percentages are calculated against the full set of reported range states, including neighbouring countries and range states not represented in ChirosetEurope.
Passive acoustic monitoring produces far more bat recordings than experts can label. We show that simple model-generated pseudo-labels turn this surplus into effective supervision. We compare pseudo-labeling with other semi-supervised learning methods on an 18-species European corpus using only 10% of its training labels, then transfer the strongest approaches to South African field audio containing nine bat taxa and a nuisance class. Pseudo-labeling outperforms the other semi-supervised learning methods on every European measure, recovering up to 61.5% of the gap to full supervision. It transfers to field audio with gains of 10.69 points in species accuracy and 4.96 points in species macro-F1. We also introduce genus-aware smoothing, which directs uncertain target mass toward congeneric species. Combined with uniform smoothing, it reaches 79.16 species macro-F1, 4.73 points above hard targets. Simple pseudo-labels are therefore highly effective at this ecological data scale, while genus-aware targets inject useful biological structure at no annotation cost. https://code4conservation.github.io/UnlabeledEchoes/
Frank Fundel, Alexandra Howard
Ludwig-Maximilians-Universität München, Germany · University of the Free State, Department of Zoology and Entomology, Qwaqwa Campus, South Africa
Animal vocalizations provide crucial insights for wildlife assessment, particularly in complex environments such as forests, aiding species identification and ecological monitoring. Recent advances in deep learning have enabled automatic species classification from their vocalizations. However, classifying species unseen during training remains challenging. To address this limitation, we introduce AnimalCLAP, a taxonomy-aware language-audio framework comprising a new dataset and model that incorporate hierarchical biological information. Specifically, our vocalization dataset consists of 4,225 hours of recordings covering 6,823 species, annotated with 22 ecological traits. The AnimalCLAP model is trained on this dataset to align audio and textual representations using taxonomic structures, improving the recognition of unseen species. We demonstrate that our proposed model effectively infers ecological and biological attributes of species directly from their vocalizations, achieving superior performance compared to CLAP. Our dataset, code, and models will be publicly available at https://dahlian00.github.io/AnimalCLAP_Page/.
Risa Shinoda, Kaede Shiohara, Nakamasa Inoue +2
The University of Osaka · The University of Tokyo · Institute of Science Tokyo
Passive acoustic monitoring holds great promise for ecological inference, yet existing automated tools are typically narrowly trained and non-transferable. We address these limitations with PULSE, a semi-supervised, multi-task framework for Orthoptera bioacoustics, combining weakly-supervised species classification, self-supervised learning on unlabelled field audio, and knowledge distillation from a general-purpose bioacoustic model. Our domain-adapted specialist model outperforms a state-of-the-art general model across all metrics (macro F1: 0.21 vs. 0.07; AUC: 0.74 vs. 0.45; AP: 0.32 vs. 0.19), with active learning further raising F1 to 0.34 and AUC to 0.84. Beyond classification, the learned embeddings encode ecologically meaningful structure, exposed through an interactive visualisation tool for ecological discovery.
Olga Isupova, Danil Kuzin, Ella Browning +2
Leverhulme Centre for Nature Recovery, School of Geography and the Environment, University of Oxford, Oxford, UK