CoChat

The DOI matches its Crossref record; 27 of the 28 references it cites were found in a registry or an index; none of the 1 sources its claims cite is checked for support yet.

Limits to the accurate and generalizable use of soundscapes to monitor biodiversity

Open access
Sethi, Sarab S., Bick, Avery, Ewers, Robert M., Klinck, Holger, Ramesh, Vijay, Tuanmu, Mao-Ning, Coomes, David A. · 2023 · Nature Ecology & Evolution

Abstract

Although eco-acoustic monitoring has the potential to deliver biodiversity insight on vast scales, existing analytical approaches behave unpredictably across studies.

Overview

CoChat Summary

TL;DR

Across 8,023 soundscape recordings paired with avifaunal point counts from four ecosystems, acoustic indices and machine learning embeddings did not generalize in predicting species richness. Soundscape change was consistently correlated with community change, but the authors conclude soundscape monitoring should be used cautiously and alongside in-person surveys.

Key findings

  • Many acoustic features correlated with avian richness within datasets, but only 4 learned features (LFs) and no soundscape indices (SSIs) correlated across all four datasets; most correlated in only one or two datasets (105 LFs, 43 SSIs).
  • Machine learning models predicted richness within datasets with R² = 0.32 (India), 0.50 (Malaysia), 0.22 (Taiwan) and 0.14 (USA), but cross-dataset performance was poor: except Taiwan-to-India (R² = 0.13), all cross-dataset R² values were < 0.
  • Pairwise soundscape change correlated with pairwise avian community change within every dataset (Spearman Mantel test, P ≤ 0.001 for all), including all 14 Malaysia sites (P ≤ 0.03), but soundscape change was not significantly correlated with change in species richness in any dataset.
  • The authors report no common acoustic features of biodiverse soundscapes and note many cases where similar soundscapes were associated with very different avian communities.

Methods

  • Collated 8,023 short (1–20 min) recordings with paired avifaunal data from four datasets: USA temperate forest (n = 6,734, one site), Malaysia tropical rainforest (n = 977, 14 sites), Taiwan tea landscape (n = 165, 16 sites) and India montane tropical forest/grassland (n = 147, 91 sites).
  • For each recording, calculated 128-dimensional VGGish CNN embeddings (LFs) and 60 analytically derived soundscape indices (SSIs); avian richness was based on unique species per point count, with eBird checklists used for the USA and expert point counts for India, Malaysia and Taiwan.
  • Used univariate Pearson correlations, random forest regression for richness prediction with 70/30 train-test splits, and Mantel tests comparing Euclidean soundscape-feature distances with Jaccard community distances.

Limitations

  • The paper states that without high-quality ground-truth data from the exact location being studied, soundscape methods should not be used to generate species-richness predictions, and that soundscape change is limited because similar soundscapes can accompany very different communities.
  • The authors suggest non-biotic sounds (for example, motors) or non-avian biotic sounds (for example, insects and amphibians) may contribute more to soundscapes than birds, so richness predictions may reflect latent habitat-suitability variables rather than direct avian vocal contributions.
  • Evident from the text: datasets differ greatly in sample size (USA n = 6,734 versus India n = 147), survey design, recording equipment and habitat, and the USA used eBird checklists rather than standardized in-person point counts; Taiwan audio was sparser and matched to point counts using a 1-hour window.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1038/s41559-023-02148-z

Cites

27 of 27 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Fig. 1 Predictions of avian species richness from acoustic features of soundscapes are not generalizable across datasets but soundscape change is a reliable indicator of community change. a, Many SSIs and LFs correlated significantly with avian species richness in each dataset. Dashed lines indicate significance thresholds derived from two-sided permutation tests and overlaps between the SSI (orange) and LF (blue) histograms are in brown. b, However, only four LFs and no SSIs correlated significantly with richness in all datasets, with most only correlating with richness in one or two datasets (105 LFs and 43 SSIs). c, For spectrogram cover and LF 14 (both of which correlated with richness in many datasets with two-sided Pearson’s tests) the gradient and intercept of the correlations varied between datasets and correlation coefficients were relatively low, limiting transferability. d, Predicting avian richness with a machine learning model trained on LFs was moderately successful when training and test datasets were the same. However, models were unable to generalize when predicting richness in datasets other than the one used for training. e, For each dataset we found that pairwise distances between the acoustic features of soundscapes correlated with Jaccard distances between avian communities present. With Jaccard distances, 0 indicates identical communities and 1 indicates no community overlap. Number of samples per dataset: India (n = 147), Malaysia (n = 977), Taiwan (n = 165) and the USA (n = 6,734).
Fig. 1 Predictions of avian species richness from acoustic features of soundscapes are not generalizable across datasets but soundscape change is a reliable indicator of community change. a, Many SSIs and LFs correlated significantly with avian species richness in each dataset. Dashed lines indicate significance thresholds derived from two-sided permutation tests and overlaps between the SSI (orange) and LF (blue) histograms are in brown. b, However, only four LFs and no SSIs correlated significantly with richness in all datasets, with most only correlating with richness in one or two datasets (105 LFs and 43 SSIs). c, For spectrogram cover and LF 14 (both of which correlated with richness in many datasets with two-sided Pearson’s tests) the gradient and intercept of the correlations varied between datasets and correlation coefficients were relatively low, limiting transferability. d, Predicting avian richness with a machine learning model trained on LFs was moderately successful when training and test datasets were the same. However, models were unable to generalize when predicting richness in datasets other than the one used for training. e, For each dataset we found that pairwise distances between the acoustic features of soundscapes correlated with Jaccard distances between avian communities present. With Jaccard distances, 0 indicates identical communities and 1 indicates no community overlap. Number of samples per dataset: India (n = 147), Malaysia (n = 977), Taiwan (n = 165) and the USA (n = 6,734).CC BY · PubMed Central
Extended Data Fig. 1 Predictions of species richness were not generalisable using alternative feature sets. When training and test datasets were the same, predictions of species richness from a machine learning model (Random Forest Regressor) trained on (a) the 60 soundscape indices (SSIs) or (b) all 188 acoustic features (60 SSIs + 128 LFs) were of varying levels of accuracy (R2 = 0.09, 0.50, 0.24, 0.10 [SSIs] R2 = 0.30, 0.54, 0.24, 0.15 [all 188 features] for India, Malaysia, Taiwan, USA, respectively). In all cases, predictive models did not generalize to produce informative estimates of species richness when applied to datasets they were not trained on. Similar results using only the 128 LFs are shown in Fig. 1c in the main text.
Extended Data Fig. 1 Predictions of species richness were not generalisable using alternative feature sets. When training and test datasets were the same, predictions of species richness from a machine learning model (Random Forest Regressor) trained on (a) the 60 soundscape indices (SSIs) or (b) all 188 acoustic features (60 SSIs + 128 LFs) were of varying levels of accuracy (R2 = 0.09, 0.50, 0.24, 0.10 [SSIs] R2 = 0.30, 0.54, 0.24, 0.15 [all 188 features] for India, Malaysia, Taiwan, USA, respectively). In all cases, predictive models did not generalize to produce informative estimates of species richness when applied to datasets they were not trained on. Similar results using only the 128 LFs are shown in Fig. 1c in the main text.CC BY · PubMed Central
Extended Data Fig. 2 Soundscape change did not correlate with change in species richness. No significant correlations were found between change in soundscape features and change in species richness in any of the four datasets (Spearman’s correlation two-sided Mantel test, p > 0.05 in all cases).
Extended Data Fig. 2 Soundscape change did not correlate with change in species richness. No significant correlations were found between change in soundscape features and change in species richness in any of the four datasets (Spearman’s correlation two-sided Mantel test, p > 0.05 in all cases).CC BY · PubMed Central
Extended Data Fig. 3 Soundscape change correlated with change in species community using soundscape indices. Even when using soundscape indices rather than learned features, change in acoustic features was correlated with change in avian community across all datasets (Spearman’s correlation two-sided Mantel test, p = 0.002, 0.001, 0.023, 0.001 for India, Malaysia, Taiwan, USA).
Extended Data Fig. 3 Soundscape change correlated with change in species community using soundscape indices. Even when using soundscape indices rather than learned features, change in acoustic features was correlated with change in avian community across all datasets (Spearman’s correlation two-sided Mantel test, p = 0.002, 0.001, 0.023, 0.001 for India, Malaysia, Taiwan, USA).CC BY · PubMed Central
Extended Data Fig. 4 Window size for the Taiwan dataset did not impact our results. The chosen size of the window around the point count start time in the Taiwan dataset did not change the accuracy or generalizability of the machine learning models trained to predict species richness.
Extended Data Fig. 4 Window size for the Taiwan dataset did not impact our results. The chosen size of the window around the point count start time in the Taiwan dataset did not change the accuracy or generalizability of the machine learning models trained to predict species richness.CC BY · PubMed Central

Verification

Claims the paper makes

0 Valid · 0 Not valid · 1 Not checked · 25 Own results and interpretations

11 more claims

Create a free account to see all 26 claims and every source verdict.

See all claims — free account
  • It may state two separate claims as one. 4
  • Probably not a substantial claim. 6

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

1 of 1

Citations

Cites

Show all 28 references

Cited by

See all 95

Metadata

Authors
Sethi, S. S., Bick, A., Ewers, R. M., Klinck, H., Ramesh, V., Tuanmu, M.-N., & Coomes, D. A.
Affiliations
  • University of Cambridge (GB)
  • Imperial College London (GB)
  • Cornell University (US)
  • Biodiversity Research Center, Academia Sinica (TW)
Corresponding author
Sethi, Sarab S.
Venue
Nature Ecology & Evolution
Published
2023-07-31
Volume
7
Issue
9
Pages
1373-1378
ISSN
2397-334X (ISSN-L)
OpenAlex
W4385416115
Cited by
95
References
28
Citations by year
2026: 35 · 2025: 29 · 2024: 25 · 2023: 6
FWCI
23.35
Type
journal article
Language
en
Open access
Licence
Locations
  • Nature Ecology & Evolution (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • Apollo (University of Cambridge) (submittedVersion, cc-by)
  • NORA - Norwegian Open Research Archives (submittedVersion, cc-by)
  • Spiral (Imperial College London) (submittedVersion, other-oa)
  • Apollo (cc-by)
Funders
  • World Wildlife Fund
  • Yayasan Sime Darby
  • Academia Sinica
  • Ministry of Science and Technology, Taiwan (MOST111-2321-B002-019)
  • National Science and Technology Council
  • Cornell Lab of Ornithology
Topics
  • Life Sciences › Agricultural and Biological Sciences › Ecology, Evolution, Behavior and Systematics › Animal Vocal Communication and Behavior
  • Physical Sciences › Environmental Science › Ecology › Wildlife Ecology and Conservation
  • Physical Sciences › Environmental Science › Ecology › Avian ecology and behavior
Keywords
soundscape, biodiversity monitoring, species richness, avifauna, machine learning prediction, community change
MeSH terms
Machine Learning, Humans, Ecosystem, Biodiversity

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Limits to the accurate and generalizable use of soundscapes to monitor biodiversity. https://app.cochat.ai/papers/10.1038/s41559-023-02148-z
The paper
Sethi, S. S., Bick, A., Ewers, R. M., Klinck, H., Ramesh, V., Tuanmu, M.-N., & Coomes, D. A. (2023). Limits to the accurate and generalizable use of soundscapes to monitor biodiversity. Nature Ecology & Evolution. https://doi.org/10.1038/s41559-023-02148-z

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.