Unsupervised source separation and clustering of soundscape recordings predicts avian richness in hyperdiverse systems
- 1Lancaster Environment Centre, Lancaster University, Lancaster, UK
Abstract
1: Passive Acoustic Monitoring (PAM) has emerged as a leading approach for biodiversity research, but analysing the vast quantities of data produced by this method remains challenging. Acoustic indices are often employed to estimate species richness in soundscape recordings, but their generalisability and overall efficacy in speciose regions has increasingly been challenged. An alternative, multi-species classification models, still suffer from performance limitations, particularly in the tropics. 2: We introduce an unsupervised approach that estimates avian species richness using an integrated source separation model trained on birds (BirdMixIT), a leading bird vocalization embedding model (Perch 2.0), and hierarchical agglomerative clustering. We count the clusters of embeddings generated from separated channels to estimate the species richness in recordings at differing timescales. 3: We test the generalisability and predictive performance of our approach, "source-clustering", using a large combined dataset featuring 103 total hours of manually-annotated soundscape recordings from the Amazon rainforest in Peru and Brazil. To compare the performance of source-clustering to existing approaches, we estimated species richness in the same recordings with both traditional acoustic indices and Perch classifications at various confidence thresholds. 4: Our source-clustering method performed consistently well between our separate datasets. Source-clustering performed better for predicting avian species richness than traditional acoustic indices at all levels, and was complementary with Perch, performing similarly overall to its best settings combinations, though it was more robust for short recordings, and should be more resilient in locations with less training data. Source-clustering generally predicted species richness better than the best-performing Perch settings at the 15-second, 1-minute, and 15-minute temporal resolutions, while most settings of Perch performed better at the 1-hour resolution. 5: Source-clustering can complement or exceed the performance of presently established methods for estimating avian species richness from soundscape recordings, namely traditional acoustic indices and the out-of-box versions of leading automated identification modes. This is particularly true at narrow temporal resolutions, in systems where training data is scarce, and where species density is high enough to generate significant overlapping vocalisations, rendering it especially useful in hyperdiverse tropical ecosystems where other methods for interpreting soundscape recordings face persistent challenges.
Related
- Concurrent Supervised-Unsupervised Representative Subspace Clustering (CSU-RSC)
- Unsupervised Clustering Enables Cohort-Specific Variant Recalibration in Crops
- Clinical Heterogeneity in Myotonic Dystrophy Type 1: Identifying Phenotype Categories Using Unsupervised Clustering.
- Decoding candidemia in critically ill patients: unsupervised clustering identifies three unique phenotypes
- Unsupervised clustering identifies candidate metabolic profiles, including an acidotic profile detectable before calving, in periparturient dairy cows
BibTeX
@misc{hopping2026unsupervised,
title = {{Unsupervised source separation and clustering of soundscape recordings predicts avian richness in hyperdiverse systems}},
author = {W. A. Hopping and S. Srinivasan and O. C. Metcalf and N. G. de Moura and A. C. Lees and E. Berenguer and L. C. Rossi and C. A. Nunes and J. Barlow},
howpublished = {bioRxiv},
year = {2026},
month = oct,
doi = {10.64898/2026.10.13.758826},
url = {https://www.biorxiv.org/content/10.64898/2026.10.13.758826v1},
note = {CC BY-NC-ND},
}
RIS
TY - UNPB
TI - Unsupervised source separation and clustering of soundscape recordings predicts avian richness in hyperdiverse systems
AU - W. A. Hopping
AU - S. Srinivasan
AU - O. C. Metcalf
AU - N. G. de Moura
AU - A. C. Lees
AU - E. Berenguer
AU - L. C. Rossi
AU - C. A. Nunes
AU - J. Barlow
PY - 2026
DA - 2026/10/13
PB - bioRxiv
DO - 10.64898/2026.10.13.758826
UR - https://www.biorxiv.org/content/10.64898/2026.10.13.758826v1
AB - 1: Passive Acoustic Monitoring (PAM) has emerged as a leading approach for biodiversity research, but analysing the vast quantities of data produced by this method remains challenging. Acoustic indices are often employed to estimate species richness in soundscape recordings, but their generalisability and overall efficacy in speciose regions has increasingly been challenged. An alternative, multi-species classification models, still suffer from performance limitations, particularly in the tropics. 2: We introduce an unsupervised approach that estimates avian species richness using an integrated source separation model trained on birds (BirdMixIT), a leading bird vocalization embedding model (Perch 2.0), and hierarchical agglomerative clustering. We count the clusters of embeddings generated from separated channels to estimate the species richness in recordings at differing timescales. 3: We test the generalisability and predictive performance of our approach, "source-clustering", using a large combined dataset featuring 103 total hours of manually-annotated soundscape recordings from the Amazon rainforest in Peru and Brazil. To compare the performance of source-clustering to existing approaches, we estimated species richness in the same recordings with both traditional acoustic indices and Perch classifications at various confidence thresholds. 4: Our source-clustering method performed consistently well between our separate datasets. Source-clustering performed better for predicting avian species richness than traditional acoustic indices at all levels, and was complementary with Perch, performing similarly overall to its best settings combinations, though it was more robust for short recordings, and should be more resilient in locations with less training data. Source-clustering generally predicted species richness better than the best-performing Perch settings at the 15-second, 1-minute, and 15-minute temporal resolutions, while most settings of Perch performed better at the 1-hour resolution. 5: Source-clustering can complement or exceed the performance of presently established methods for estimating avian species richness from soundscape recordings, namely traditional acoustic indices and the out-of-box versions of leading automated identification modes. This is particularly true at narrow temporal resolutions, in systems where training data is scarce, and where species density is high enough to generate significant overlapping vocalisations, rendering it especially useful in hyperdiverse tropical ecosystems where other methods for interpreting soundscape recordings face persistent challenges.
ER -