Source-linked AI summary
Effects of interpulse-interval variation on deep-learning classification of bat vocalizations
Welmoed R. Eversteijn, Burooj Ghani, A. Leonie Baier, Dan Stowell
TL;DR
The paper tests whether interpulse-interval variation contributes to bat-species classification and whether transformer models use it more effectively than CNNs. Across matched natural- and normalized-timing evaluations, evidence for a dominant IPI contribution was limited, while normalized-training results did not fully transfer to natural recordings.
Problem
Call features vary across species, habitats, and behaviors, creating overlap; whether IPI sequence context helps classification and whether transformers use it more effectively than CNNs remained unclear.
Source: §1 · pp. 1–2Method
The study compared matched European bat recordings with natural or normalized call timing, fine-tuning EfficientNet-B0 and PaSST and evaluating pretrained BatDetect2 and BAT across conditions.
Source: §1 · p. 2, §2.1 · p. 2, §2.3 · p. 4Results
Takeaways & Limitations
Normalized call spacing may simplify within-condition classification without necessarily improving performance on unmodified recordings; evaluation should retain natural variability expected during deployment.
Source: §5 · p. 10The EfficientNet–PaSST comparison cannot isolate architecture effects because their pretraining, input representations, spectrogram parameters, temporal resolutions, and preprocessing also differed.
Source: §4.3 · p. 9
Abstract
from arXiv · showhide
Temporal context may aid automated bat-species classification, but the contribution of specific features remains unclear. We investigated whether variation in the interpulse interval (IPI)-the time between consecutive call onsets-provides species-discriminative information and whether transformer-based models are more sensitive to this information than convolutional neural networks. We created two matched datasets from European bat recordings: a natural-IPI condition retaining the original call timing and a normalized-IPI condition in which call onsets were spaced at 50-ms intervals. EfficientNet-B0 and PaSST were fine-tuned and evaluated within each condition. In an additional experiment, each architecture was trained separately on natural-IPI and normalized-IPI recordings, and evaluated on the same natural-IPI test set. Finally, the pretrained classifiers BatDetect2 and BAT were evaluated on both conditions. Within-condition IPI normalization had model-dependent effects. PaSST accuracy differed little between the natural-IPI ($71 \pm 2.3\%$) and normalized-IPI ($70 \pm 6.3\%$) conditions, whereas EfficientNet accuracy increased from $47 \pm 4.7\%$ to $57 \pm 3.9\%$. PaSST exceeded EfficientNet under both conditions. In the cross-condition evaluation, models trained on natural-IPI recordings outperformed those trained on normalized-IPI recordings on the natural-IPI test set: accuracy decreased from 54% to 50% for EfficientNet and from 65% to 57% for PaSST. BatDetect2 and BAT differed little between IPI conditions. Overall, we found limited support for the hypotheses that natural IPI variation contributes substantially to bat-species classification and that it is used more effectively by transformer-based than CNN-based models. Nevertheless, the cross-condition performance decrease shows that results obtained under normalized conditions may not transfer fully to natural recordings.
1. Introduction
Affordable large-scale monitoring is needed as biodiversity declines, and passive acoustic monitoring makes continuous bat recording feasible. Because call structure varies with behavior and overlaps across species, the study asks whether IPI variation provides useful sequence context for classification.
1. Introduction
- Passive acoustic monitoring uses autonomous sensors to record environmental sounds continuously, supporting more feasible large-scale monitoring as sensors become less expensive.It addresses limits of handheld detectors, which provide snapshots that do not accurately portray biodiversity.Source: §1 · p. 1
- Call duration, bandwidth, and frequency modulation vary with species, habitat, and behavior, creating overlap between species’ acoustic features.Calls typically become shorter and more broadband as bats move from search to approach and capture.Source: §1 · p. 1
- IPI variation may help a classifier locate an individual call within a behavioral sequence and distinguish species despite overlapping call features.The study tests whether transformers use this temporal context more effectively than CNNs.Source: §1 · p. 2
- The study isolates call timing by comparing matched datasets that retain natural spacing or normalize intervals while preserving individual pulses’ spectral content.The authors investigate whether IPI variation contributes to automated classification and whether transformers leverage it more than CNNs.Source: §1 · p. 2
2. Data and methods
The study compares bat classifiers using matched recordings that retain natural call timing or reduce variation in call-onset intervals. It combines within-condition cross-validation, a held-out cross-condition test, and evaluation of existing pretrained classifiers.
2.1. Study design and objectives
- Experiment 1 compared two pretrained pipelines initialized from EfficientNet-B0 and PaSST, fine-tuning all layers; five-fold stratified cross-validation was conducted within each natural-IPI and normalized-IPI condition.Source: §2.1 · p. 2, §2.2 · p. 3
- Three experiments compare within-condition performance, generalization from normalized to natural timing, and existing classifiers’ sensitivity to IPI variation.Natural IPI preserves original temporal spacing; normalized IPI reduces variation between call onsets.Source: §2.1 · p. 2, Table 1 · p. 3
- In Experiment 2, models trained separately on natural-IPI or normalized-IPI recordings were evaluated on the same held-out natural-IPI test set.Matched versions of each source recording were kept in the same partition.Source: §2.1 · p. 2, §2.2 · p. 3
2.2. Data collection and dataset composition
- Matched datasets were created from European bat recordings, with call timing either retained or normalized while the conditions were intended to differ only in temporal structure.The normalized condition set intervals between call onsets to 50 ms.Source: §2.2 · p. 2, §2.1 · p. 2
3. Existing-model evaluation
The methods compare existing bat classifiers on natural and normalized recordings and describe how data were annotated, processed, trained, and evaluated. They also specify the cross-validation and held-out designs used for the fine-tuned models.
2.3. IPI annotation and normalization
- IPI normalization aligned call onsets to a regular 50-ms sequence anchored at the first annotated onset, while reducing variation between consecutive calls.The target interval was chosen to exceed expected pulse durations and reduce overlap risk; manual onset uncertainty may leave small deviations.Source: §2.3 · pp. 3–4
2.4. Model preprocessing, configuration and implementation
- EfficientNet-B0 used resized, normalized linear spectrograms, whereas PaSST used mel spectrograms after tenfold time expansion and resampling.Both models were fully fine-tuned; configurations and training procedures are described in the paper’s model setup.Source: §2.4 · p. 4
2.5. Experiment 1: within-condition training and cross-validation
- Experiment 1 used five-fold stratified cross-validation within each condition, shared fold assignments, and validation macro-F1 for hyperparameter selection and early stopping.No independent test set was used.Source: §2.5 · p. 5
2.6. Experiment 2: cross-condition held-out evaluation
- Experiment 2 compared separate natural-IPI and normalized-IPI training runs on one fixed natural-IPI test set, with matched source recordings kept in the same partition.Because no separate validation set was used, models trained for 50 epochs and the lowest-training-loss checkpoint was evaluated.Source: §2.6 · p. 5
3. Existing-model evaluation
- Existing pretrained classifiers were evaluated without additional training on natural-IPI and normalized-IPI recordings, with predictions converted to recording-level species labels.Evaluation covered the complete dataset and model-specific subsets containing species in each classifier’s original training taxonomy.Source: §2.7 · p. 5
2.7. Experiment 3: existing pretrained models
- BatDetect2 and BAT predictions were reduced to recording-level labels using each model’s most frequent detected species or highest-ranked species, respectively.Predicted labels, rather than probabilities, were compared with reference species labels.Source: §2.7 · p. 5
2.8. Evaluation metrics
- Performance was assessed with accuracy and macro-averaged precision, recall, and F1-score, plus per-species metrics to examine variation among species.Macro-averaging gave each species equal weight because the dataset remained imbalanced.Source: §2.8 · p. 5
3. Results
Across matched-condition evaluation, PaSST outperformed EfficientNet, but IPI normalization affected the models differently; training on normalized-IPI recordings also reduced accuracy on a natural-IPI test set. Existing pretrained classifiers changed little across IPI conditions, and their complete-dataset accuracy was low.
3.1. Dataset composition and IPI normalization
- The matched datasets contained 251 two-second segments per condition across 16 species, with natural call timing or call onsets fixed at 50 ms.The cross-validation and held-out experiments used 201 training and 50 validation or test segments, respectively.Source: §3.1 · pp. 5–6, Figure 2 · p. 6
3.2. Experiment 1: within-condition model training and cross-validation
- PaSST accuracy was 71% ± 2.3% on natural IPI and 70% ± 6.3% on normalized IPI, while EfficientNet rose from 47% ± 4.7% to 57% ± 3.9%.PaSST outperformed EfficientNet under both conditions.Source: §3.2 · p. 6, Figure 2 · p. 6
3.3. Experiment 2: cross-condition held-out
- On the natural-IPI test set, training on normalized rather than natural IPI reduced accuracy from 65% to 57% for PaSST and from 54% to 50% for EfficientNet.PaSST remained more accurate than EfficientNet.Source: §3.3 · p. 6, Figure 2 · p. 6
3.4. Experiment 3: existing pretrained models
- On complete datasets, BatDetect2 and BAT each scored 28% natural-IPI and 27% normalized-IPI accuracy; known-species subsets scored 58%/56% and 38%/36%, respectively.The known-species values are natural-IPI/normalized-IPI accuracy for BatDetect2 and then BAT.Source: §3.4 · p. 6, Figure 2 · p. 6
4. Discussion
Across experiments, evidence that natural IPI variation was a dominant species cue was limited, and the results do not establish that transformers use temporal context more effectively than CNNs. Normalization gains were condition-dependent and did not fully transfer to natural recordings.
4.1. Sensitivity to IPI variation
- BatDetect2 and BAT changed little across IPI conditions, while normalized-IPI training improved within-condition performance for both fine-tuned models, especially EfficientNet-B0.The proposed explanation that normalization reduced unrelated temporal variability is tentative, and other effects of the procedure cannot be excluded.Source: §4 · p. 8, §4.1 · p. 8
4.2. Generalization to natural recordings
- Models trained on natural-IPI recordings performed better on the natural-IPI test set; normalized-condition gains did not fully transfer, and evidence for dominant IPI contribution remained limited.The cross-condition result does not establish that models learned IPI as a species-specific feature.Source: §4.2 · p. 8, §4 · p. 8
- The cross-condition estimates are uncertain: each species had approximately four test recordings, no separate validation set was available, and estimates may depend on test-set composition.Hyperparameter selection and training duration also warrant consideration when interpreting the results.Source: §4.2 · p. 8
- Broader generalization remains unestablished because labels may be uncertain, clips capture limited variation, and the dataset duplicates serotine labels, potentially affecting aggregate metrics.The findings therefore require confirmation with larger, more diverse datasets.Source: §4.2 · pp. 8–9
4.3. Comparison of model pipelines
- Neither model comparison isolates architecture: the fine-tuned systems differ in pipeline components, and BatDetect2 and BAT also differ in datasets, species, outputs, and prediction procedures.The combined results leave open whether CNNs or transformers are generally better suited to bat-species classification.Source: §4.3 · p. 9
4.4. Implications
- The findings concern IPI spacing within two-second recordings, not temporal context generally; call order, morphology changes, rhythm, and longer-term behavior were not independently manipulated.PaSST's higher performance does not by itself show that it used temporal information other than IPI.Source: §4.4 · p. 9, §4.3 · p. 9
- Regularizing call timing can simplify classification without improving natural-recording performance, so evaluations using similarly processed training and test data may overestimate field performance.The authors recommend representative natural variability and cross-condition tests for passive acoustic monitoring.Source: §4.4 · p. 9
- The timing-manipulation framework can support targeted ablations; future work could use larger, diverse datasets and independently manipulate IPI, call order, morphology, and background structure.Matched model inputs, pretraining, and optimization would strengthen CNN–transformer comparisons.Source: §4.4 · p. 9
5. Conclusion
The findings provide limited evidence that natural IPI variation is a dominant source of species-discriminative information or that transformers use it more effectively than CNNs. Normalized-condition results did not transfer fully to natural recordings, and broader data and targeted manipulations are needed to clarify sequence-context effects.
5. Conclusion
- The results provide limited evidence that natural IPI variation is a dominant source of species-discriminative information or is used more effectively by transformers than CNNs.BatDetect2 and BAT changed little between IPI conditions, and IPI reduction did not affect PaSST more strongly than EfficientNet.Source: §5 · p. 9
- Normalized-IPI training reduced accuracy on the natural-IPI test set for both architectures, but the decrease cannot be attributed specifically to IPI variation.The evaluation indicates sensitivity to training–test condition differences, while leaving open whether other changes from normalization contributed.Source: §5 · pp. 9–10
- Larger, more diverse datasets and targeted temporal and spectral manipulations are needed to clarify sequence-context contributions and model differences.The findings also support evaluating classifiers on data that retain natural variability expected during deployment.Source: §5 · p. 10
Funding
The study received support through the Horizon Europe TETTRIs and MAMBO projects; funders had no role in the research or its reporting.
Funding
- Funding came through Horizon Europe projects TETTRIs and MAMBO, and the funders had no role in study design, data collection and analysis, publication decisions, or manuscript preparation.The authors state that the views are theirs and do not necessarily reflect those of the European Union or REA.Source: Funding · p. 10