A Parameter-Free Few-Shot Evaluation for Elephant Vocalisation Classification

Christiaan M. Geldenhuys; Thomas R. Niesler

Source-linked AI summary

A Parameter-Free Few-Shot Evaluation for Elephant Vocalisation Classification

Christiaan M. Geldenhuys, Thomas R. Niesler

arXiv:2608.14824v1Aug 14, 2026eess.AScs.LGcs.SDq-bio.QM

TL;DR

Elephant vocalisation classification often has few labelled exemplars, making it important to assess simple classifiers in low-resource settings. This paper evaluates nearest-centroid classification on fixed pretrained embeddings as exemplars per class vary, finding it competitive with trained baselines on the small EV dataset but not the larger LDC dataset.

Problem
Evidence is limited on how the simplest classifier performs as labelled exemplars per class vary, despite scarce annotations constraining elephant call classification.
Method
The paper evaluates parameter-free nearest-centroid classification on fixed pretrained acoustic embeddings in an N-way k-shot setting across EV and LDC datasets.
Results
On EV, centroid classification with Perch embeddings surpasses trained baselines from one or two exemplars, whereas on LDC it remains below them across the few-shot regime.
Takeaways & Limitations
With a suitable pretrained embedding and few labelled exemplars, centroid classification can be deployed without training, validation, or tuning stages.
Takeaways & Limitations
Centroid performance is limited by whether the embedding encodes discriminative features for the relevant taxa and call types.

Abstract

from arXiv · show

We present a parameter-free episodic evaluation of nearest-centroid classification for elephant vocalisations on fixed pretrained acoustic embeddings, across the Elephant Voices (EV) and Linguistic Data Consortium (LDC) datasets. Rather than asking which embedding yields the best classifier when trained on all available labelled data, we ask how the simplest classifier performs as labelled exemplars per class are varied. Each class is represented by the mean of its support-set embeddings, and each query is assigned to the nearest centroid under squared Euclidean distance. We evaluate this centroid classifier on the Perch (ver. 1), Perch (ver. 2), and HuBERT (base, layer 2) embeddings, together with mel frequency cepstral coefficient (MFCC) features, in an N-way k-shot manner under the same cross-validation protocol as the trained baselines. A bootstrap over 100 resampled support sets quantifies the sampling noise. On the smaller, low-resource EV dataset, the centroid classifier using the stronger Perch (ver. 1) and Perch (ver. 2) embeddings overtakes the fully-trained logistic regression classifier from a single exemplar per class and the stronger recurrent classifier from two. Over the reduced set of call types on which the strongly-supervised end-to-end baseline was trained, the centroid classifier matches and then surpasses that baseline in mean average precision (mAP), from a few exemplars per class. On the larger LDC dataset, where labelled exemplars are abundant, the trained baselines retain their advantage at every k considered. At five exemplars per class, the centroid classifier using the strongest embedding, Perch (ver. 2), attains a mAP of 0.542 on the EV dataset and 0.368 on the LDC dataset. Parameter-free nearest-centroid classification is the stronger choice when labelled exemplars are few and the fixed embedding already encodes the features that separate the call types.

1 Introduction

Automated elephant-vocalisation classification could support conservation monitoring, but scarce and noisy expert labels limit deeper neural approaches. This work therefore evaluates a parameter-free nearest-centroid classifier on fixed embeddings as labelled exemplars per class vary, comparing low- and high-data regimes and support-set sampling uncertainty.

Automated classification could enable continuous passive monitoring, inform wildlife management, and provide early warning of conservation threats such as poaching.
Scarce, expensive, and noisy expert annotations have limited deeper neural-network approaches for elephant call classification.Noise can arise from boundary disagreements, interference, overlapping vocalisations, and inter-annotator variability.
The study varies labelled exemplars per class for a no-parameter nearest-centroid classifier operating on fixed pretrained embeddings.Each query is assigned to the nearest class mean of its support set, making the approach feasible when labels are scarce.
The episodic evaluation covers EV and LDC datasets, Perch (ver. 1), Perch (ver. 2), HuBERT, and MFCC features using the baselines’ stratified cross-validation folds.Every episode spans all C evaluated call types.
The study compares parameter-free and trained systems across contrasting labelled-data regimes and quantifies support-set sampling noise with bootstrap percentile intervals.It also examines unequal exemplars per class, distinguishing complete from partial support sets and identifying when interval narrowing reflects exhausted labelled data.

2 Background

Elephant call classification is constrained by scarce labelled recordings, motivating few-shot evaluation with fixed acoustic embeddings. The section defines episodic nearest-centroid classification as a parameter-free alternative to trained shallow classifiers.

Motivation: Scarce labelled recordings limit neural approaches to elephant call classification and make classifier training, model selection, and hyperparameter tuning impractical.The challenge is especially acute when labelled exemplars are very scarce.
Acoustic embeddings: The study evaluates Perch (ver. 1), Perch (ver. 2), and HuBERT (base, layer 2) embeddings, with MFCC features as a comparative baseline.Acoustic embeddings are fixed, high-dimensional numerical representations of audio signals.
Few-shot learning: Few-shot learning classifies new inputs from only a small number of labelled exemplars per class, using a pretrained embedding model as a fixed feature extractor.A shallow classifier such as logistic regression can be trained directly on the resulting representations.
Episodic evaluation: Each episodic task samples N classes and k labelled exemplars per class into a support set, while a disjoint query set is withheld for scoring.The support set contains m = N × k exemplars, and limited labelled data may require sampling with replacement.
Nearest-centroid classifier: The parameter-free nearest-centroid classifier averages support embeddings by class and assigns each query to the closest centroid under squared Euclidean distance.It contains no learnable parameters, so episodic evaluation does not require training or development sets.

3 Literature Review

Research on elephant vocalisation analysis has progressed from early MFCC- and HMM-based classification to modern deep-learning, transfer-learning, and monitoring systems. Parallel few-shot bioacoustic work established standardised protocols, prototypical-network methods, and active-learning approaches for sparse labels.

Elephant acoustic monitoring and classification: Elephant vocalisation research spans more than two decades, beginning with HMMs and MFCCs for call-type classification and individual-caller identification.Clemins et al. (2005) presented one of the earliest machine-learning approaches in this area.
Elephant acoustic monitoring and classification: Recent elephant-monitoring research includes cross-modal machine-learning reviews, CNN-based vocalisation detection, and transfer-learning workflows using pretrained acoustic models.The cited transfer-learning workflow fine-tunes VGGish, YAMNet, Perch, and BirdNet for forest elephant monitoring.
Few-shot bioacoustics with fixed embeddings: The DCASE few-shot bioacoustic challenge formalised prototypical-network detection from as few as five labelled examples and standardised evaluation across species and recording conditions.The challenge also highlighted the difficulty of generalisation across species and recording conditions.
Few-shot bioacoustics with fixed embeddings: Few-shot active learning combines human-in-the-loop annotation, pretrained AST support embeddings, and boundary-focused sample selection to detect rare or sparse animal vocalisations.Samples closest to the decision boundary, representing the lowest-certainty cases, are prioritised for annotation.
Few-shot bioacoustics with fixed embeddings: On few-shot whale-call classification, Perch (ver. 2) outperformed BirdNet and Perch (ver. 1) in marine-mammal evaluation.Burns et al. (2025) evaluated Perch (ver. 2) on marine mammal tasks.

4 Datasets

The study evaluates two elephant-vocalisation datasets separately because they differ in size, recording setup, and constituent species. EV contains fine-grained subcall annotations from African elephants, whereas LDC contains coarser call-type annotations from Asian elephants.

Dataset overview: The study uses the Elephant Voices (EV) and Linguistic Data Consortium (LDC) datasets, evaluated separately because of differences in size, recording setup, and constituent species.The datasets are summarised separately before segmentation, with evaluated segments reported later.
Elephant Voices dataset: 55.6 minutes of EV field recordings contain vocalisations of free-roaming African bush elephants.EV annotations are provided at recording level, so multi-call recordings were divided into shorter single-call segments.
Linguistic Data Consortium dataset: 57.4 hours of LDC field recordings, including 32.6 annotated hours, contain vocalisations of free-roaming Asian elephants.The corpus defines 16 categories: 14 vocal call types plus the blow and trunk-bounce non-vocal acoustic signals.
Annotation taxonomy: EV distinguishes fine-grained subcall types, while LDC distinguishes coarser call types, and each dataset is evaluated at its annotation level.Examples include the greeting rumble and musth rumble in EV, versus rumble and trumpet in LDC.
Preprocessing and evaluation: All recordings are resampled to 16 kHz, converted to single channel, and peak-normalised to −3 dB relative to full scale.The same sample-level stratified cross-validation folds are used for both datasets.

5 Experimental Setup … 5.3 Cross-validated episode construction

The evaluation uses fixed pretrained acoustic embeddings with parameter-free nearest-centroid classification, based on human-annotated call segments and cross-validated few-shot episodes. It applies the same fold structure as supervised baselines while varying the support exemplars available to each class.

5.1 Audio segmentation: The procedure segments long recordings into annotated vocalisation intervals, using known human-provided endpoints rather than automatic endpointing.Because vocalisations may overlap, the task is fundamentally multi-label, although each segment is treated as a single-label sample during evaluation.
5.1 Audio segmentation: The evaluated data comprise 422 annotated EV subcall-type segments and 4845 LDC call segments.EV segmentation produced 1063 segments, including 594 call segments; only the 422 with subcall-type labels were used.
5.1 Audio segmentation: The common fold is held out from cross-validation and never queried, but its segments remain eligible as support exemplars.It contains material that cannot be distributed across all K folds, principally call types occurring fewer than K times.
5.2 Embedding models: All models remain fixed without weight updates or fine-tuning, supplying embeddings to both nearest-centroid classification and the logistic-regression baseline.The evaluated representations are Perch (ver. 1), Perch (ver. 2), HuBERT, and MFCC features.
5.2 Embedding models: MFCC representations average per-frame vectors computed with 25 ms frames, 10 ms stride, a 128-band mel filterbank, and 40 retained coefficients.HuBERT uses averaged 768-dimensional frame embeddings from its base variant.
5.3 Cross-validated episode construction: Cross-validation uses five stratified, recording-grouped folds for EV and ten for LDC, with labels taken from each segment’s temporal centre.At inference, a prediction is correct if it matches any human-annotated label associated with the segment.
5.3 Cross-validated episode construction: The centroid classifier reuses the supervised baselines’ folds, drawing queries from the held-out test fold and supports from the remaining training folds.Each fold serves as the query fold in turn, so every segment outside the common fold is evaluated once as a query.

5.4 Nearest-centroid classifier · 5.5 Varying the number of support exemplars · 5.6 Baseline classification systems

The evaluation uses a parameter-free nearest-centroid classifier whose representations are recomputed from each support set, varying the number of labelled exemplars per class against fully supervised baselines. Baselines include linear logistic regression, recurrent classifiers, and the strongly supervised AERD end-to-end system.

5.4 Nearest-centroid classifier: Each class centroid is computed directly from its support set, and queries are assigned using squared Euclidean distance.The classifier has no learnable parameters, trained weights, validation set, early stopping, or tuned hyperparameters.
5.4 Nearest-centroid classifier: A new class centroid is computed in every evaluation episode from the associated support set, with no state carried between episodes.
5.5 Varying the number of support exemplars: The experiment characterises centroid-classification performance as the number of labelled support exemplars per class, k, increases from one exemplar toward using all eligible support-fold exemplars.At the upper limit, the support set contains all labelled data available to a class and represents the fully supervised training regime.
5.5 Varying the number of support exemplars: A class has a complete support set when it provides at least k distinct labelled exemplars, and a partial support set when it provides fewer than k.
5.6 Baseline classification systems: The AERD system is a strongly supervised AST-seq end-to-end baseline for elephant call classification, detection, and automatic endpointing.It is pretrained on AudioSet using a self-supervised learning approach and adapted for sequence-to-sequence modelling.
5.6 Baseline classification systems: Logistic regression provides a linear baseline whose performance indicates class separability in the fixed embedding space.For each embedding, the study also reports the best cross-validated recurrent classifier: a GRU, LSTM, or Elman RNN, depending on embedding and dataset.
5.6 Baseline classification systems: Within each outer cross-validation fold, logistic regression and recurrent baselines use all labelled training data, whereas the centroid classifier uses only k support exemplars per class.This intentional asymmetry gives the trained baselines full access to the training fold.

5.7 Bootstrapped estimation of support set sampling noise

The procedure estimates classification-performance variability caused by randomly sampled few-shot support sets while holding query segments fixed. It precomputes support-dependent scores, resamples them into bootstrap datasets, and summarizes sampling noise using percentile intervals.

Sampling-noise source: Support-set sampling makes classification performance a random variable because query segments remain fixed while centroid-forming exemplars are drawn randomly.This isolates sampling noise from variation in the query segments.
Score precomputation: The resulting Q × R × N score array is computed once, allowing bootstrap samples to be assembled without further inference.This avoids repeating the costly centroid formation and query-to-centroid distance calculations.
Score precomputation: R = 100 independently sampled support sets are drawn with replacement for each query, producing support-dependent score vectors for every query–draw pair.Each support set yields class centroids and softmax scores for the candidate classes.
Bootstrap evaluation: B = 5000 bootstrap samples independently select one precomputed score vector per query and evaluate either mAP or AUROC within folds, averaged across folds.The procedure is metric-agnostic and applies to any performance summary.
Noise summary: The mean and 1st-to-99th-percentile interval of bootstrap performance values summarize classification performance and its support-set sampling noise.The interval width represents the spread induced by support-set sampling.

5.8 Metrics

The study reports AUROC and mAP as threshold-independent classifier-performance metrics. AUROC summarizes ROC performance across classes, while average precision measures ranking quality through the precision-recall curve.

Reported metrics: AUROC and mAP are threshold-independent performance summaries, unlike fixed-threshold measures such as accuracy.The two reported metrics are area under the receiver operating characteristic curve and mean average precision.
AUROC: AUROC aggregates true-positive and false-positive rates across decision thresholds and is informative under class imbalance because it is not biased by class prior.For the multi-class setting, AUROC is computed per call type and averaged without weighting.
Mean average precision: Average precision is the area beneath a class-specific precision-recall curve formed by ranking query segments by their classifier scores.The area depends on precision at ranking positions where segments belonging to that class appear.

6 Results

Nearest-centroid performance improves with more exemplars, with Perch embeddings strongest across EV and LDC. It surpasses trained baselines in parts of the low-resource EV setting but remains below them on LDC.

EV dataset: Perch (ver. 2) centroid mAP on EV rises from 0.385 at k = 1 to 0.521 at k = 4, with diminishing successive gains.At k = 1, it already exceeds the Perch (ver. 2) LR baseline of mAP = 0.365.
LDC dataset: On LDC, the centroid classifier remains below fully supervised baselines through the few-shot regime and reaches the Perch (ver. 2) LR baseline only near k = 35.At k = 40, it reaches mAP = 0.487 versus the LR baseline mAP = 0.482; in AUROC, it remains below at 0.905 against 0.932.
EV dataset: On EV, Perch (ver. 2) and Perch (ver. 1) centroid classifiers exceed their corresponding LR and recurrent baselines at larger k.The embedding ordering is Perch (ver. 2), Perch (ver. 1), HuBERT (base, layer 2), then MFCC.

7 Discussion

The discussion frames nearest-centroid classification as the same linear model as logistic regression, with parameters fixed from support-set means rather than optimized. Its advantage depends on the bias–variance trade-off, embedding quality, and whether additional exemplars remain diverse.

7.1 Linear-classifier interpretation: Nearest-centroid classification has the same linear form as logistic regression, but sets weights and biases from class centroids rather than iterative optimization.Centroid parameters come from a handful of support exemplars, whereas logistic regression optimizes against all labelled training data.
7.1 Linear-classifier interpretation: The centroid rule imposes isotropic within-class covariance, reducing estimation variance but limiting adaptation, whereas trained weights remove this bias while increasing variance with few exemplars.These strategies therefore occupy opposite ends of the bias–variance trade-off.
7.2 Performance plateau: On EV, Perch (ver. 2) and Perch (ver. 1) centroid performance plateaus at mAP ≈0.59, above the LR and recurrent baselines, while MFCC and HuBERT plateau lower.The weaker-feature plateaus are approximately 0.35 for MFCC and 0.52 for HuBERT (base, layer 2).
7.2 Performance plateau: Centroid gains diminish because sample-mean variance falls roughly as 1/k, while beyond k⋆= 4 on EV, partial classes exhaust their eligible support pools.On LDC, every class remains complete through k⋆= 31, so support exhaustion affects only larger k values beyond the few-shot regime.
7.2 Performance plateau: The embedding sets the performance ceiling: centroid classification exceeds LR and Elman RNN baselines on EV but remains below LR and GRU baselines on LDC.For call types absent from embedding pretraining, more support exemplars are unlikely to help if the representation does not encode discriminative features.

8 Summary and Conclusion

The parameter-free nearest-centroid classifier is especially data-efficient on the small EV dataset when paired with stronger pretrained embeddings, surpassing trained baselines with very few labelled exemplars. This supports deployment without classifier training while making embedding coverage the main remaining limitation.

Method: The evaluation tested nearest-centroid classification on fixed pretrained acoustic embeddings while varying labelled exemplars per class.Each class was represented by the mean of its support-set embeddings, and queries were assigned to the nearest centroid.
Results: On the small EV dataset, stronger Perch embeddings made centroid classification notably data-efficient.Using Perch (ver. 1) and Perch (ver. 2), performance matched and then surpassed fully-trained LR from one exemplar per class and surpassed the stronger recurrent classifier from two.
Practical implication: With a handful of labelled exemplars per call type, a centroid classifier can be deployed without training, validation, or tuning.This directly addresses scarce-label tasks such as subcall classification for conservation practitioners.
Limitation: The embedding itself remains the principal limitation, making expanded pretraining for uncovered taxa and call types the likely route to further gains.This shifts future improvement toward broader pretrained embedding coverage rather than classifier complexity.

ORCID

The paper lists ORCID identifiers for Christiaan M. Geldenhuys and Thomas R. Niesler.

Christiaan M. Geldenhuys is identified by ORCID 0000-0003-0691-0235, while Thomas R. Niesler is identified by ORCID 0000-0002-7341-1017.

Appendix A Evaluated call types

The appendix defines the evaluated call types at each dataset’s annotation level and identifies the largest complete shot count for each type. It also distinguishes EV’s two non-vocal acoustic signals from its vocal taxonomy and specifies the four coarser AERD categories used for centroid scoring.

Evaluated call types: Call types are evaluated at each dataset’s annotation level: EV uses subcall types, whereas LDC uses coarser call types.The evaluated types are ordered by the number of labelled exemplars.
Evaluated call types: The final column reports the largest shot count k for which each call type remains complete across every cross-validation split.A call type is complete when it can supply k support exemplars in every split.
Evaluated call types: The blow and trunk-bounce are non-vocal acoustic signals outside the 14-call vocal taxonomy and do not belong to an AERD category.These signals are attributed to de Silva (2010).
Evaluated call types: Centroid classification is scored over four coarser AERD categories: rumble, roar, bark, and trumpet.The categories are used in Table 6b.