Pith. sign in

REVIEW 3 major objections 5 minor 88 references

Adding metadata as auxiliary training targets improves species identification and transfer across acoustic, species, and geographic domain shifts.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 02:50 UTC pith:37K7GQKS

load-bearing objection A solid, honest empirical study whose central claim—location, season, and background-species auxiliary losses improve bioacoustic transfer—is plausible and mostly supported, but the noise in its model-selection procedure makes the reported gains less certain than the text suggests. the 3 major comments →

arxiv 2607.14072 v1 pith:37K7GQKS submitted 2026-07-15 cs.LG cs.SD

MetaPerch: Learning from metadata for bioacoustics foundation models

Mustafa Chasmai , Vincent Dumoulin , Jenny Hamer This is my paper
classification cs.LG cs.SD
keywords bioacousticsfoundation modelsmetadata as auxiliary supervisionmulti-task learningspecies identificationdomain shiftpassive acoustic monitoringlocation prediction
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to show that recording metadata—chiefly location, season, and background species—can serve as auxiliary supervision for a bioacoustic foundation model, improving species identification when the model is later tested on soundscapes, novel species, and new regions. A same-architecture baseline trained only on species labels is compared against MetaPerch, which adds small prediction heads for metadata during training only; MetaPerch reports gains of +0.015 ROC-AUC on BirdSet, +0.016 accuracy and +0.006 cmAP on BEANS, and +0.018 prototype / +0.070 linear ROC-AUC on WABAD. The sympathetic reader cares because deployment in passive acoustic monitoring requires transfer under exactly these shifts, and metadata is freely available in citizen-science training data but not needed at test time.

Core claim

The central claim is that forcing the shared audio embedding to predict metadata, as auxiliary tasks alongside species identification, produces features that transfer better under domain shift. MetaPerch jointly trains species classification with prediction of location (S2 cell at level 7), season, and background species, treating each as a classification task with a constant loss weight and with mixup targets built from whichever metadata is present. Missing metadata is handled by placeholder values and zeroed loss contributions; an optional gradient-reversal layer makes the embedding adversarial to a metadata source. On the paper's own terms, the discovery is that this simple multi-task re

What carries the argument

The central object is a multi-task training objective on a shared EfficientNet-B3 embedding: the standard species classification loss plus a weighted sum of metadata classification losses, each with its own small MLP head. Location is binned into S2 cells (roughly 50,000 square kilometers at the chosen level) and treated as a classification target; season is derived from date and hemisphere; background species is a multi-label target present on a fraction of recordings. The mechanism does its work by shaping the embedding to encode where and when a recording was made and what else is audible, so that at test time—where metadata is not supplied—the embedding already carries geographic and hab

Load-bearing premise

The load-bearing premise is that the validation recipe—geometric averaging of observed-species classification, one-shot retrieval, and novel-species linear probing—predicts performance on the 17 evaluation benchmarks, so that the choices of location/season/background metadata, loss weights, and hyperparameters are supported; the paper itself reports a disconnect where the median hyperparameter trial beats the best-validation trial on WABAD.

What would settle it

Train MetaPerch and the species-only baseline with identical compute but select hyperparameters using a held-out portion of WABAD (or another globally distributed soundscape benchmark) instead of the paper's three validation task families. If the WABAD gains of +0.018 prototype / +0.070 linear ROC-AUC shrink or vanish, the reported benefit is largely a model-selection artifact rather than a property of metadata supervision. A second check: remove the location head and confirm whether the WABAD biome-clustering visualization collapses; if it does not, location supervision is not the driver of t

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Metadata supervision only needs to be available during training; deployed models can run without location or time inputs.
  • Location and season help most for regions underrepresented in training data, such as South America and Hawai‘i in BirdSet.
  • The same recipe with the same architecture outperforms the species-only baseline on soundscape tasks and on novel-species transfer, by 0.015–0.070 ROC-AUC depending on benchmark.
  • Background species supervision appears to teach the embedding to attend to co-occurring species, which is exactly the signal needed in dense soundscape recordings.
  • Simple constant loss weights are sufficient; more complex multi-task balancing strategies do not add value.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the mechanism is really about encoding habitat and geographic priors, the same auxiliary heads could be grafted onto other audio foundation models or taxa; a direct test would be to train location and season heads on a general audio encoder and measure transfer on non-avian tasks.
  • The paper's own validation-disconnect result suggests the reported gains may understate or overstate deployment benefit depending on how models are selected; a fair comparison would select hyperparameters with WABAD-style globally distributed soundscapes in the validation mix.
  • Because even 1% metadata availability still improved over baseline, metadata supervision could be a practical pretraining signal for passive acoustic monitoring deployments where species labels are sparse but GPS and timestamps are routinely logged.
  • Adversarial training helped for day-part and text-embedding metadata, hinting that a per-metadata adversarial-versus-standard switch could be automated rather than chosen by hand.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper introduces MetaPerch, a bioacoustic foundation model trained on Xeno-Canto, iNaturalist, Tierstimmenarchiv, and FSD50K. In addition to the species-identification objective, MetaPerch adds auxiliary prediction heads for recording metadata — most importantly location (S2-cell classification), season, and background species — with per-task loss weights tuned by Vizier. The model is compared to BioBaseline, a same-architecture model without metadata losses, and to recent baselines on BirdSet (7 datasets), BEANS (10 bioacoustic tasks), WABAD, and controlled BIRB few-shot settings. The paper reports average improvements of +0.015 BirdSet ROC-AUC, +0.016 accuracy/+0.006 cmAP on BEANS, and +0.018/+0.070 ROC-AUC on WABAD, along with extensive ablations of metadata sources, missing-metadata handling, mixup variants, adversarial training, and location formulations.

Significance. Metadata as auxiliary supervision is an attractive, low-cost direction: metadata is already present in citizen-science repositories and can be used at training time without test-time conditioning. If the reported gains are robust, MetaPerch would offer a simple recipe for improving transfer under acoustic, species, and geographic shift. The paper's controlled same-architecture baseline, the breadth of the evaluation (17 datasets), the systematic ablation of nine metadata sources, and the 5-seed repeated runs are genuine strengths, as is the public model release. The central caveat is model selection: the paper reports a disconnect between its validation score and WABAD performance, and the hyperparameter space for MetaPerch is much larger than for BioBaseline, so the size of the improvement may be inflated by selection. This is addressable with additional analysis, but it currently tempers confidence.

major comments (3)
  1. [§3.4, §4.4 (Table 5a)] The model-selection procedure in §3.4 chooses the metadata source set, per-metadata loss weights, head architectures, and adversarial/standard flags by a geometric average over three validation task families. Section 4.4 then reports that the median Vizier trial by validation score yields 0.952 WABAD ROC-AUC, while the best-validation trial yields only 0.946. Thus the validation ranking is not predictive of WABAD, one of the headline benchmarks. Because MetaPerch has many more tunable choices than BioBaseline, selecting the single best-validation configuration creates a best-of-N selection bias; the 5-seed standard deviations in §A.7 quantify seed noise for the selected configuration, not selection variance. The reported gains are therefore not yet established as unbiased estimates of the metadata benefit. Please report the full distribution of validation/test scores (or a selection-adju
  2. [Table A6] The appendix reports Top-1 accuracy for BirdSet. MetaPerch is worse than BioBaseline on four of seven datasets (UHH 0.550 vs 0.651; HSN 0.580 vs 0.625; NBP 0.679 vs 0.717; SNE 0.778 vs 0.792) and better only on PER and SSW. The main text's characterization of 'consistent improvements' rests entirely on ROC-AUC. Since the paper's central claim is that metadata improves species identification, this metric sensitivity needs to be discussed directly: either justify ROC-AUC/cmAP as the primary competencies (and state this in the main text), or explain why Top-1 accuracy regresses. As written, a reader could reasonably conclude that the metadata model trades classification confidence for ranking quality.
  3. [§4.4, Figure 7a] The BIRB few-shot analysis shows that MetaPerch's advantage over BioBaseline reverses at higher shot counts on the Xeno-Canto held-out set. The paper acknowledges this and appeals to PAM soundscapes being more representative. This is a legitimate qualification, but the reversal is also present in a species-shift-only setting, which is central to the 'novel species' narrative. Please either provide an explanation for the reversal or soften the general claim in the abstract/conclusion that metadata improves transfer to novel species. At minimum, the limitation should be reflected in the summary rather than only in a figure caption.
minor comments (5)
  1. [§3.3] Typo: 'biacoustics' should be 'bioacoustics'.
  2. [Table A3] The hidden-dimension search space lists 265 where 256 is clearly intended (it appears as '265{64, 128, 265, 512}' and the optimal value is also 265). Please correct the typo.
  3. [Table A6] '0.58±.014' should be '0.580±.014' for consistency with the other entries.
  4. [Figure 7a] The legend contains duplicated/confusing entries ('MetaBio', 'BioBaseline' twice). Please clean up the figure legend.
  5. [Abstract/Conclusion] The abstract and conclusion claim that metadata improves species identification quite broadly. The reported WABAD desert biome drop (Table A8, Deserts/Xeric Shrublands: 0.936 to 0.912) and the BIRB reversal at higher shots are counterexamples that appear only in the body; a sentence of qualification would better reflect the evidence.

Circularity Check

0 steps flagged

No derivation-level circularity: metadata improvements are empirical, held-out benchmark results; self-citations and validation selection do not reduce to the paper's own inputs.

full rationale

MetaPerch's central claim is empirical: adding location, season, and background-species auxiliary losses improves species identification relative to a same-architecture BioBaseline on BirdSet, BEANS, and WABAD. This is not derived from the training objective by construction; it is measured on held-out evaluation benchmarks. The paper's self-citations (Perch 2.0, BIRB, Chasmai et al. 2026) supply architecture, baselines, and experiment templates, not the metadata-improvement result, and the comparisons also include external methods (Perch 2.0, BirdNet, etc.). The validation/model-selection procedure in Section 3.4 uses a geometric average over validation tasks that do not overlap the test benchmarks; the paper explicitly discloses a 'disconnect between the validation and WABAD performance' (Section 4.4, Table 5a), where the median Vizier trial outperforms the best-validation trial on WABAD. This is a disclosed selection-validity limitation, not a circular reduction: the final WABAD or BirdSet numbers are not forced by the validation scores, and the paper does not hide the discrepancy. Ablations in Section 4.4 use WABAD to compare design choices (e.g., location formulation in Figure 7d), which is a test-set selection risk, but it does not make the reported gains equivalent to the selected hyperparameters by construction. Overall, no step in the claimed derivation chain reduces to its own inputs, so no circularity is identified.

Axiom & Free-Parameter Ledger

10 free parameters · 6 axioms · 0 invented entities

The central claim rests on a large empirical training setup rather than a mathematical derivation. The main free choices are the metadata source set, the S2 location level, and metadata loss weights/hyperparameters, all selected on validation tasks. The key assumptions are that citizen-science metadata are accurate and ecological, weak labels from recordings are valid, and evaluation sets are disjoint from training.

free parameters (10)
  • Metadata loss weights (lambda_location, lambda_season, lambda_background) = 1.006, 0.955, 1.042 (Table A3)
    Tuned by Vizier on validation; they determine the balance of auxiliary supervision in Eq. 6 and directly affect the central claim.
  • Metadata source set = {location, season, background species} of 9 candidates
    Selected after model selection (Sec. 3.3, 4.4); the central claim depends on which metadata are chosen.
  • S2 cell level for location = Level 7 (~50,000 km^2)
    Chosen from the formulations in Figure 7d; coarser or finer levels change location task difficulty and information content.
  • Metadata head architectures = location: swish/265/1; season: relu/512/2; background: relu/512/1 (Table A3)
    Vizier hyperparameters; head capacity affects whether auxiliary tasks are learnable and how much they shape the shared embedding.
  • Adversarial-vs-standard training flags = False (standard) for all selected metadata
    Vizier explored both options (Sec. 3.3.1); choosing standard over adversarial training changes how spurious correlations are treated.
  • Learning rate = 2.98e-4
    Vizier-tuned optimization parameter (Table A3).
  • Dropout = 0.389
    Vizier-tuned regularization parameter (Table A3).
  • Mixup parameters = Dirichlet concentration 5.73, max signals 5, beta-binomial (15.09, 12.84)
    Augmentation distributions tuned via Vizier; the metadata mixing rule is also chosen by validation (Table 5b).
  • Season and day-part category definitions = 4 classes each; broad boundaries
    Hand-defined discretizations in A.4; different bucketing changes the auxiliary task difficulty and semantics.
  • Background-species missing-metadata/mixup rule = 'mix if any present'
    Chosen over 'mix iff all' and weighted-mean variants (Table 5b); affects the amount of label noise in metadata supervision.
axioms (6)
  • domain assumption All labels of a recording are assigned to every sampled 5s window (weak labels).
    A.3: annotations lack time bounds; this standard assumption can inject label noise but is needed for training on citizen-science data.
  • domain assumption Cross-dataset taxonomy mapping (Xeno-Canto/Tierstimmenarchiv to iNaturalist classes) is correct.
    A.1; mapping errors would corrupt class labels and metadata correlations.
  • domain assumption Citizen-science metadata (location, date, background species) are accurate and ecologically meaningful, not just platform artifacts.
    Sec. 3.3.1 discusses adversarial training because 'some species-metadata correlations may reflect training biases'; the selected model assumes the correlations are useful.
  • domain assumption Evaluation benchmarks are disjoint from training data; only CBI is explicitly excluded for overlap.
    Sec. 4.2 and Table 3 caption; if other benchmarks overlap with training, gains could be inflated.
  • ad hoc to paper Mixup with multihot metadata targets ('mix any present metadata') is valid supervision.
    Sec. 3.3.1; chosen by validation over stricter and weighted variants (Table 5b).
  • standard math Standard spectrogram construction and S2 geometry are correct.
    A.5/A.6: STFT/log-mel spectrograms and S2 cell computations are taken as given.

pith-pipeline@v1.3.0-alltime-deepseek · 24409 in / 14716 out tokens · 138943 ms · 2026-08-02T02:50:04.787620+00:00 · methodology

0 comments
read the original abstract

Bioacoustic foundation models rely on large-scale citizen science platforms like Xeno-Canto for geographically and ecologically diverse data. Recent work has shown that supervision alone can produce SotA species detection models when trained on this large-scale data -- however, there remains unutilized potential in the form of recording metadata readily available within these community-driven data hubs. In this work, we explore the use of metadata -- such as location and time -- as auxiliary supervision signals, allowing the model to leverage species-metadata correlations in its learned representation. Auxiliary metadata losses provide additional information beyond vocalizations alone that can encourage a richer, more robust representation that generalizes better to species distribution and acoustic domain shifts -- important challenges for deployment in real-world passive acoustic monitoring (PAM) settings. We introduce MetaPerch, a new foundation model that achieves strong species identification performance across multiple challenging domains and present an extensive empirical study of the effects of 9 diverse metadata sources on 17 bioacoustic datasets.

Figures

Figures reproduced from arXiv: 2607.14072 by Jenny Hamer, Mustafa Chasmai, Vincent Dumoulin.

Figure 1
Figure 1. Figure 1: Problem setup. Bioacoustics training and evalua￾tion settings differ significantly. This figure illustrates substantial acoustic (focal → soundscape) and species (observed → novel) domain shifts in our setup. In this work, we explore the use of metadata as auxiliary supervision to learn robust, generalizable features and help bridge the domain gap. marks (Rauch et al., 2025b; Hagiwara et al., 2023; Chasmai… view at source ↗
Figure 2
Figure 2. Figure 2: Dataset statistics. Geographical distribution of the training (log scale) and evaluation (rough locations) datasets. tics and present training recipes for metadata-aware species identification. Metadata outside bioacoustics. The use of metadata, par￾ticularly location, has received significant interest in other fields as well. Soundscape mapping (Khanal et al., 2024) aims to understand the acoustic signatu… view at source ↗
Figure 3
Figure 3. Figure 3: Method diagram. Overview of BioBaseline and METAPERCH. Location Season Day Part BackgroundQuality Samp. Rate Call Type Life Stage Sex MicrophoneNotes 0 20 40 60 80 100 Percentage of recordings Combined as free-form text [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Metadata availability. Percentage of recordings across the full training dataset which have each metadata type. main shift. In such cases, we can encourage embeddings to be agnostic to—rather than predictive of—such “spurious” metadata by leveraging an idea introduced in domain ad￾versarial networks (Ganin et al., 2016) and using a gradient reversal layer (Osumi et al., 2019) between the backbone and the m… view at source ↗
Figure 5
Figure 5. Figure 5: Performance on bioacoustic benchmarks. Visualiza￾tion of our performance relative to the baseline. For each dataset, we present the same metric used in Tables 2 and 3 (ROC-AUC for BirdSet, Accuracy & cmAP for BEANS). We highlight the domain shifts (species, acoustic, geographic) in each dataset to better illustrate scenarios where metadata is beneficial. those consisting of PAM soundscapes (with dense labe… view at source ↗
Figure 6
Figure 6. Figure 6: Embedding visualizations. 2D UMAP projections of BioBaseline (left) and METAPERCH (right) embeddings of WABAD windows, colored by biome. 4.4. Ablations METAPERCH design choices. First, we study the ef￾fects of different design choices in METAPERCH. For the multi-loss training strategy (Table 5a), our validation per￾formance suggests that treating the weights of each meta￾data loss as hyperparameters to be … view at source ↗
Figure 7
Figure 7. Figure 7: Ablations. For (a), we report few-shot linear probe performance on BIRB (Hamer et al., 2023). For (b-d), we report ROC￾AUC achieved by prototype learning head on WABAD. they are available for only a fraction of the training data. Individual metadata performance. We observe improve￾ments by encouraging features to be predictive of most metadata on their own, with boosts ranging from 0.008 for Day-Part to 0.… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

88 extracted references · 1 canonical work pages

  1. [1]

    European conference on computer vision , pages=

    Planet-photo geolocation with convolutional neural networks , author=. European conference on computer vision , pages=. 2016 , organization=

    2016

  2. [2]

    2021 , publisher=

    Kahl, Stefan and Wood, Connor M and Eibl, Maximilian and Klinck, Holger , journal=. 2021 , publisher=

    2021

  3. [3]

    Scientific Reports , volume=

    Global birdsong embeddings enable superior transfer learning for bioacoustic classification , author=. Scientific Reports , volume=. 2023 , publisher=

    2023

  4. [4]

    arXiv preprint arXiv:2312.07439 , year=

    Hamer, Jenny and Triantafillou, Eleni and Van Merri. arXiv preprint arXiv:2312.07439 , year=

    Pith/arXiv arXiv

  5. [5]

    arXiv preprint arXiv:2508.11845 , year=

    What matters for bioacoustic encoding , author=. arXiv preprint arXiv:2508.11845 , year=

    Pith/arXiv arXiv

  6. [6]

    arXiv preprint arXiv:2508.01277 , year=

    Foundation models for bioacoustics---a comparative review , author=. arXiv preprint arXiv:2508.01277 , year=

    arXiv

  7. [7]

    arXiv preprint arXiv:2508.04665 , year=

    Perch 2.0: the bittern lesson for bioacoustics , author=. arXiv preprint arXiv:2508.04665 , year=

    arXiv

  8. [8]

    EfficientNet : Rethinking model scaling for convolutional Neural Networks

    Tan, Mingxing and Le, Quoc. EfficientNet : Rethinking model scaling for convolutional Neural Networks. International Conference on Machine Learning

  9. [9]

    A closer look at loss weighting in multi-task learning , author=

  10. [10]

    Advances in neural information processing systems , volume=

    Gradient surgery for multi-task learning , author=. Advances in neural information processing systems , volume=

  11. [11]

    Forty-third International Conference on Machine Learning , year=

    Bioacoustic Geolocation: Species Sounds as Geographic Signals , author=. Forty-third International Conference on Machine Learning , year=

  12. [12]

    International Conference on Learning Representations , year=

    mixup: Beyond Empirical Risk Minimization , author=. International Conference on Learning Representations , year=

  13. [13]

    2019 16th International Conference on Machine Vision Applications (MVA) , pages=

    Domain adaptation using a gradient reversal layer with instance weighting , author=. 2019 16th International Conference on Machine Vision Applications (MVA) , pages=. 2019 , organization=

    2019

  14. [14]

    Journal of machine learning research , volume=

    Domain-adversarial training of neural networks , author=. Journal of machine learning research , volume=

  15. [15]

    iNaturalist Research-grade Observations

    iNaturalist contributors. iNaturalist Research-grade Observations

  16. [16]

    Xeno-Canto

    Vellinga, Willem-Pier and Planqu\' e , Bob. Xeno-Canto

  17. [17]

    The archive of animal sounds at the Humboldt-university of Berlin

    Frommolt, Karl-Heinz. The archive of animal sounds at the Humboldt-university of Berlin. Bioacoustics

  18. [18]

    FSD50K : An open dataset of human-labeled sound events

    Fonseca, Eduardo and Favory, Xavier and Pons, Jordi and Font, Frederic and Serra, Xavier. FSD50K : An open dataset of human-labeled sound events. IEEE/ACM Transactions on Audio, Speech, and Language Processing

  19. [19]

    Advances in Neural Information Processing Systems , volume=

    The iNaturalist sounds dataset , author=. Advances in Neural Information Processing Systems , volume=

  20. [20]

    Symes and Viviana Ruiz-Gutiérrez and Ingrid Molina-Mora and Fernando Cediel and Luis Sandoval and Holger Klinck , title =

    Álvaro Vega-Hidalgo and Stefan Kahl and Laurel B. Symes and Viviana Ruiz-Gutiérrez and Ingrid Molina-Mora and Fernando Cediel and Luis Sandoval and Holger Klinck , title =. Zenodo , version = 1, doi =

  21. [21]

    Zenodo , version = 1, doi =

    Mary Clapp and Stefan Kahl and Erik Meyer and Megan McKenna and Holger Klinck and Gail Patricelli , title =. Zenodo , version = 1, doi =

  22. [22]

    Alexander Hopping and Stefan Kahl and Holger Klinck , title =

    W. Alexander Hopping and Stefan Kahl and Holger Klinck , title =. Zenodo , version = 1, doi =

  23. [23]

    Wood and Philip Chaon and M

    Stefan Kahl and Connor M. Wood and Philip Chaon and M. Zachariah Peery and Holger Klinck , title =. Zenodo , version = 1, doi =

  24. [24]

    Zenodo , version = 1, doi =

    Amanda Navine and Stefan Kahl and Ann Tanimoto-Johnson and Holger Klinck and Patrick Hart , title =. Zenodo , version = 1, doi =

  25. [25]

    Zenodo , version = 1, doi =

    Stefan Kahl and Russell Charif and Holger Klinck , title =. Zenodo , version = 1, doi =

  26. [26]

    PeerJ Computer Science , volume=

    NIPS4Bplus: a richly annotated birdsong audio dataset , author=. PeerJ Computer Science , volume=. 2019 , publisher=

    2019

  27. [27]

    Marín-Gómez and Irene Mendoza and Miguel A

    Pérez-Granados, Cristian and Jon Morant and Kevin Darras and Oscar H. Marín-Gómez and Irene Mendoza and Miguel A. Muñoz-Mohedano and Eduardo Santamaría-García and Giulia Bastianelli and Alba Márquez-Rodríguez and Michał Budka and Gerard Bota and José M. De la Peña-Rubio and Eladio L. García de la Morena and Manu Santa-Cruz and Pablo de la Nava and Mario F...

  28. [28]

    Weldy and Tom Denton and Abram B

    Matthew J. Weldy and Tom Denton and Abram B. Fleishman and Jaclyn Tolchin and Matthew Mckown and Robert S. Spaan and Zachary J. Ruff and Julianna M. A. Jenkins and Matthew G. Betts and Damon B. Lesmeister , title =. Zenodo , version =. doi:10.5281/zenodo.10895837 , url =

    work page doi:10.5281/zenodo.10895837

  29. [29]

    Scientific Data , volume=

    A Public Dataset of Annotated Orcinus orca Acoustic Signals for Detection and Ecotype Classification , author=. Scientific Data , volume=. 2025 , publisher=

    2025

  30. [30]

    Pacific Islands Passive Acoustic Network ( PIPAN ) 10kHz Data

    NOAA Pacific Islands Fisheries Science Center. Pacific Islands Passive Acoustic Network ( PIPAN ) 10kHz Data

  31. [31]

    Philosophical Transactions B , volume=

    Using tropical reef, bird and unrelated sounds for superior transfer learning in marine bioacoustics , author=. Philosophical Transactions B , volume=. 2025 , publisher=

    2025

  32. [32]

    Proceedings of Meetings on Acoustics , volume=

    The Watkins marine mammal sound database: an online, freely accessible resource , author=. Proceedings of Meetings on Acoustics , volume=. 2016 , organization=

    2016

  33. [33]

    Scientific data , volume=

    An annotated dataset of Egyptian fruit bat vocalizations across varying contexts and during vocal ontogeny , author=. Scientific data , volume=. 2017 , publisher=

    2017

  34. [34]

    2020 , howpublished =

    Addison Howard and Holger Klinck and Sohier Dane and Stefan Kahl and tom denton and Tom Denton , title =. 2020 , howpublished =

    2020

  35. [35]

    Animal behaviour , volume=

    Barking in domestic dogs: context specificity and individual identification , author=. Animal behaviour , volume=. 2004 , publisher=

    2004

  36. [36]

    Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2) , year=

    HumBugDB: A Large-scale Acoustic Mosquito Dataset , author=. Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2) , year=

  37. [37]

    , author=

    Few-Shot Bioacoustic Event Detection: A New Task at the DCASE 2021 Challenge. , author=. DCASE , pages=

    2021

  38. [38]

    , author=

    An annotated set of audio recordings of Eastern North American birds containing frequency, time, and species information. , author=. Ecology , volume=

  39. [39]

    Hawaiian Islands cetacean and ecosystem assessment survey (HICEAS) towed array data. edited and annotated for the 9th international workshop on detection, classification, localization, and density estimation of marine mammals using passive acoustics (DCLDE 2022) , author=. 2022 , publisher=

    2022

  40. [40]

    Ecological Informatics , volume=

    A pipeline for identification of bird and frog species in tropical soundscape recordings using a convolutional neural network , author=. Ecological Informatics , volume=. 2020 , publisher=

    2020

  41. [41]

    Remote Sensing in Ecology and Conservation , year=

    Automated detection of Hainan gibbon calls for passive acoustic monitoring , author=. Remote Sensing in Ecology and Conservation , year=

  42. [42]

    Advances in Neural Information Processing Systems , volume=

    Geoclip: Clip-inspired alignment between locations and images for effective worldwide geo-localization , author=. Advances in Neural Information Processing Systems , volume=

  43. [43]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Satclip: Global, general-purpose location embeddings with satellite imagery , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  44. [44]

    The Thirteenth International Conference on Learning Representations , year=

    BirdSet: A Large-Scale Dataset for Audio Classification in Avian Bioacoustics , author=. The Thirteenth International Conference on Learning Representations , year=

  45. [45]

    ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=

    Beans: The benchmark of animal sounds , author=. ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=. 2023 , organization=

    2023

  46. [46]

    arXiv preprint arXiv:2411.07186 , year=

    Naturelm-audio: an audio-language foundation model for bioacoustics , author=. arXiv preprint arXiv:2411.07186 , year=

    Pith/arXiv arXiv

  47. [47]

    The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

    Towards a translative model of Sperm Whale vocalization , author=. The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

  48. [48]

    arXiv preprint arXiv:2504.12880 , year=

    Can Masked Autoencoders Also Listen to Birds? , author=. arXiv preprint arXiv:2504.12880 , year=

    Pith/arXiv arXiv

  49. [49]

    arXiv preprint arXiv:2605.14031 , year=

    Masked Autoencoders with Limited Data: Does It Work? A Fine-Grained Bioacoustics Case Study , author=. arXiv preprint arXiv:2605.14031 , year=

    Pith/arXiv arXiv

  50. [50]

    The Thirty-Ninth Annual Conference on Neural Information Processing Systems workshop: AI for non-human animal communication , year=

    TinyBird: Fine-Grained Syllable Discovery in Birdsong Using Asymmetric Patches , author=. The Thirty-Ninth Annual Conference on Neural Information Processing Systems workshop: AI for non-human animal communication , year=

  51. [51]

    ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=

    Aves: Animal vocalization encoder based on self-supervision , author=. ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=. 2023 , organization=

    2023

  52. [52]

    arXiv preprint arXiv:2406.01253 , year=

    animal2vec and MeerKAT: A self-supervised transformer for rare-event raw audio input and a large-scale reference dataset for bioacoustics , author=. arXiv preprint arXiv:2406.01253 , year=

    Pith/arXiv arXiv

  53. [53]

    George Vengrovski and Timothy J Gardner , booktitle=. Song. 2025 , url=

    2025

  54. [54]

    International conference on machine learning , pages=

    Spatial implicit neural representations for global-scale species mapping , author=. International conference on machine learning , pages=. 2023 , organization=

    2023

  55. [55]

    ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=

    Exploring meta information for audio-based zero-shot bird classification , author=. ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=. 2024 , organization=

    2024

  56. [56]

    Ecology letters , volume=

    AVONET: morphological, ecological and geographical data for all birds , author=. Ecology letters , volume=. 2022 , publisher=

    2022

  57. [57]

    Global ecology and biogeography , volume=

    Life-history characteristics of European birds , author=. Global ecology and biogeography , volume=. 2018 , publisher=

    2018

  58. [58]

    ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=

    Transferable models for bioacoustics with human language supervision , author=. ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=. 2024 , organization=

    2024

  59. [59]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Presence-only geographical priors for fine-grained image classification , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  60. [60]

    Ecological Informatics , volume=

    Improving deep learning acoustic classifiers with contextual information for wildlife monitoring , author=. Ecological Informatics , volume=. 2023 , publisher=

    2023

  61. [61]

    Expert Systems with Applications , volume=

    DeepShip: An underwater acoustic benchmark dataset and a separable convolution based autoencoder for classification , author=. Expert Systems with Applications , volume=. 2021 , publisher=

    2021

  62. [62]

    The Thirty-Ninth Annual Conference on Neural Information Processing Systems workshop: AI for non-human animal communication , year=

    Perch 2.0 transfers 'whale' to underwater tasks , author=. The Thirty-Ninth Annual Conference on Neural Information Processing Systems workshop: AI for non-human animal communication , year=

  63. [63]

    arXiv preprint arXiv:2108.09203 , year=

    Parsing birdsong with deep audio embeddings , author=. arXiv preprint arXiv:2108.09203 , year=

    Pith/arXiv arXiv

  64. [64]

    Interspeech 2024 , pages=

    Investigating self-supervised speech models' ability to classify animal vocalizations: The case of gibbon's vocal signatures , author=. Interspeech 2024 , pages=. 2024 , organization=

    2024

  65. [65]

    BioRxiv , pages=

    TweetyNet: a neural network that enables high-throughput, automated annotation of birdsong , author=. BioRxiv , pages=. 2020 , publisher=

    2020

  66. [66]

    ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=

    Acoustic identification of individual animals with hierarchical contrastive learning , author=. ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=. 2025 , organization=

    2025

  67. [67]

    ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=

    Comparing Self-Supervised Learning Models Pre-Trained on Human Speech and Animal Vocalizations for Bioacoustics Processing , author=. ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=. 2025 , organization=

    2025

  68. [68]

    Methods in Ecology and Evolution , volume=

    Same data, different results? Machine learning approaches in bioacoustics , author=. Methods in Ecology and Evolution , volume=. 2025 , publisher=

    2025

  69. [69]

    arXiv preprint arXiv:2409.08589 , year=

    Domain-invariant representation learning of bird sounds , author=. arXiv preprint arXiv:2409.08589 , year=

    arXiv

  70. [70]

    Ecological Informatics , volume=

    Unsupervised classification to improve the quality of a bird song recording dataset , author=. Ecological Informatics , volume=. 2023 , publisher=

    2023

  71. [71]

    Proceedings of the 2024 International Conference on Information Technology for Social Good , pages=

    Active and Transfer Learning for Efficient Identification of Species in Multi-Label Bioacoustic Datasets , author=. Proceedings of the 2024 International Conference on Information Technology for Social Good , pages=

    2024

  72. [72]

    Proceedings of the IEEE/CVF winter conference on applications of computer vision , pages=

    Multi-loss weighting with coefficient of variations , author=. Proceedings of the IEEE/CVF winter conference on applications of computer vision , pages=

  73. [73]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Multi-task learning using uncertainty to weigh losses for scene geometry and semantics , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  74. [74]

    Proceedings of the International Conference on Learning Representations (ICLR) , year=

    Gradient Vaccine: Investigating and Improving Multi-task Optimization in Massively Multilingual Models , author=. Proceedings of the International Conference on Learning Representations (ICLR) , year=

  75. [75]

    Advances in neural information processing systems , volume=

    Multi-task learning as multi-objective optimization , author=. Advances in neural information processing systems , volume=

  76. [76]

    arXiv preprint arXiv:2409.07825 , year=

    Deep multimodal learning with missing modality: A survey , author=. arXiv preprint arXiv:2409.07825 , year=

    Pith/arXiv arXiv

  77. [77]

    ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=

    Improving bird classification with unsupervised sound separation , author=. ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=. 2022 , organization=

    2022

  78. [78]

    Advances in neural information processing systems , volume=

    This looks like that: deep learning for interpretable image recognition , author=. Advances in neural information processing systems , volume=

  79. [79]

    Ecological Informatics , volume=

    AudioProtoPNet: An interpretable deep learning model for bird sound classification , author=. Ecological Informatics , volume=. 2025 , publisher=

    2025

  80. [80]

    2025 , author =

    S2 Geometry Library , howpublished =. 2025 , author =

    2025

Showing first 80 references.