henhaobio
Sign in

Ecological InformaticsEcologyPeer-reviewed journal article

One-Class Bioacoustic Detector for Monitoring the Critically Endangered Pied Tamarin (Saguinus bicolor)

Juan Gabriel ColonnaiD, Tainara Venturini SobrozaiD, Marcelo GordoiD, Eduardo Freire NakamuraiD, Alejandro C. FreryiD

PDFDOIPreprintCodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

The pied tamarin (Saguinus bicolor) is a critically endangered primate with a small geographic range that includes fragmented urban forest mosaics in Amazonia, where habitat subdivision and anthropogenic actions complicate its survival and monitoring. Passive acoustic monitoring (PAM) offers a convenient, noninvasive way to track this species, yet open-set rainforest soundscapes make single-species detection challenging. We present a machine-learning pipeline with a very low false-positive rate, appropriate for downstream inference. The method combines a band-pass filter (5 - 10 kHz), Perch bioacoustic embeddings (deep learning), and a One-Class SVM (OCSVM) applied to sliding windows of continuous audio recordings to detect S. bicolor calls. We train on a reduced dataset of labeled calls and validate against diverse out-of-class audio (birds, anurans, anthropophony, and geophony/insects), then test on long, cross-site recordings. The approach achieves high discrimination on held-out negatives and produces very low false-positive rate in continuous, real-world audio, with a precision of 0.86. Finally, we pair detections with a single-site occupancy model in a cross-site setting to illustrate end-to-end utility for conservation monitoring and to estimate the false-negative detection probability in recordings from pied tamarin populations in a different geographic region. Our strategy provides a tool for PAM of S. bicolor that requires minimal manual labeling effort and can be adapted to other open-set, single-species monitoring scenarios. We grant reproducibility by releasing a Python package (https://pypi.org/project/sauim-detector/), installable via 'pip', that processes an audio file and produces detection timestamps as an Audacity label file (.txt), enabling faster manual verification.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 4 (28 Aug 2026) · CC BY 4.0

Band-pass-filtered Perch embeddings scored by a One-Class SVM trained on 82 pied tamarin calls detect the species cross-site with precision 0.86 but limited sensitivity.

Monitoring the critically endangered pied tamarin by passive acoustics is an open-set problem: Amazonian soundscapes hold an effectively unbounded, site-shifting set of non-target sounds that a binary classifier cannot represent, and existing detectors such as BirdNET do not target this species. This work fits a one-class boundary around target calls only, using pretrained Perch embeddings so that a small labeled set suffices.

Key results

  1. SuggestedA pipeline of band-pass filter, Perch embeddings and OCSVM, trained only on target calls, separated held-out S. bicolor calls from five distinct negative sound sets in a simulated open-set test, showing that a bird-trained embedding transfers to a primate.

    OCSVM fit on 82 calls with ν tuned by LOOCV; test set of 36 held-out calls and 54 background segments, plus bird, anuran, anthropophony and geophony+insect negative sets; weighted F1 stayed above 97% for all five. Authors warn the small held-out set can yield near-perfect metrics. · 3.1. S. bicolor and background noises dataset; 4. Experimental setup; 5.1. Classification

  2. ShownRestricting input to the 5 to 10 kHz band where tamarin call energy concentrates improved separability, both on the held-out test set and in cross-site continuous audio.

    On the Background test set, filtering cut false positives from 4 to 0 and false negatives from 6 to 3; ROC curves shifted toward higher AUC with lower threshold sensitivity; on the Mindú Park recording, filtering raised AUC and F1 while precision stayed high (Table 2). · 5.1. Classification; 6. Use case: a detection example; Figure legends; Tables

  3. ShownOn a ten-minute recording from Mindú Park, a site and tamarin group not used in training, the detector was highly specific but missed most annotated calls, labeling segments positive only when confident.

    5 s windows with a one-second hop (80% overlap) gave 559 segments and 51 detections; precision 0.86 on continuous audio; event-level FNR 69% and FPR 0.03% (Section 6), restated as FPR = 0.03, FNR = 0.69, TPR = 0.31, TNR = 0.97 (Section 9.2). Low F1 came from low recall at high precision. · Abstract; 6. Use case: a detection example; 9.2. Specificity, sensitivity, and the role of…

  4. SuggestedTreating the Mindú Park recording days as repeated visits to one occupied site, a single-season occupancy likelihood turns detector outputs into a visit-level detection probability.

    With Ψ assumed 1, R = 1 and T = 8, the binomial MLE gave p11 = y/T = 0.75, so visit-level miss probability p01 = 0.25; authors note this differs from the event-level FNR of 0.69 because each visit offers many detection chances. p10 needs unoccupied sites and was not estimated. · 7. Cross-site validation of the OCSVM detector…; 9.4. From acoustic detections to ecological…

  5. ShownAn MIT-licensed, pip-installable Python package (sauim-detector) runs the full pipeline locally and outputs detection timestamps for Audacity, plus a filtered audio file, so experts review only candidate segments.

    Given an audio file, the package applies band-pass filtering, Perch embedding and OCSVM scoring, then writes an Audacity-importable .txt label file and a band-pass-filtered .wav (Figure 9 shows the label track over a spectrogram). No runtime or throughput figures are given. · Abstract; 8. Software package; Figure legends

Methods, data and reproducibility

Recordings are resampled to 32 kHz, passed through a zero-phase fourth-order Butterworth band-pass (5 to 10 kHz), cut into 5 s windows, embedded with Perch (1280-dim), and scored by an RBF One-Class SVM trained only on target calls with ν tuned by LOOCV for AUC. Day-level detections then feed a single-site occupancy model.

  • Pied tamarin recordings from Bosque da Ciência (INPA); source of 118 call and 179 background 5 s segments used for training and hold-out testing · Saguinus bicolor (pied tamarin) · n = 32 recordings (about 2 h 17 min)
  • Pied tamarin recordings from Mindú Park, a separate group not in contact with the INPA group; used for cross-site detection and the eight-day occupancy example · Saguinus bicolor (pied tamarin) · n = 38 recordings (about 2 h 38 min)
  • Birds negative set: Xeno-canto recordings from Manaus and surroundings · birds (26 species) · n = 30 recordings; 130 five-second segments
  • Anuran negative set (frogs and toads), some co-occurring with the pied tamarin · anurans (ten species) · n = 60 recordings; 122 five-second segments
  • Anthropophony negative set assembled from ESC-50 clips (aircraft, traffic, chainsaws, sirens, fireworks, engines, etc.) · n = 84 five-second segments
  • Geophony and insect negative set from SoundCloud (rain, wind, thunder, fire, waves, crickets, flies) · n = 77 five-second segments
Methods & tools
Perch bioacoustic embedding model (EfficientNet-B1, about 7.8 M parameters), One-Class SVM with RBF kernel (scikit-learn), Fourth-order Butterworth band-pass filter (SciPy sosfiltfilt), Leave-one-out cross-validation for ν selection, Principal Component Analysis, Hierarchical clustering (cosine distance, average linkage), Single-season occupancy model (MacKenzie et al. 2003; Royle and Link 2006), sauim-detector Python package, Audacity, ESC-50 corpus, Xeno-canto repository, Sennheiser ME67 microphone with H4N recorder
Software & environment
Python with scikit-learn (OCSVM) and SciPy (sosfiltfilt); Perch model from Kaggle; ready-to-run Jupyter notebooks; sauim-detector installable via pip under MIT license. No package versions or containers given.
Reproducibility gaps
No software versions, hardware or runtimes are stated; the Perch embedding model is obtained from an external Kaggle repository.
Limitations (6)
  • The held-out test set is small (36 calls) relative to training, and scarce S. bicolor data can produce near-perfect classification metrics, so hold-out figures may overstate real performance. · stated by the authors
  • Event-level sensitivity across sites is low (FNR 0.69), calls with two or fewer bouts raise misses, and onset annotations may misalign with analysis windows, so the detector should not serve alone to infer absence. · stated by the authors
  • The negative sets give no systematic test against Saguinus midas, a sympatric tamarin whose calls may shift toward the pied tamarin pattern, so species-level discrimination where the two overlap remains unverified. · stated by the authors
  • The one-class boundary reflects only the call variability in the training group, and Mindú recordings were made close to the animals; other populations or fixed recorders farther away (lower SNR) may degrade transfer and detection-probability estimates. · stated by the authors
  • No alternative detector (e.g., BirdNET, a binary classifier, or other embeddings without the one-class step) is benchmarked on the same data, so the gain from the one-class formulation is not quantified. · noted in this summary
  • Cross-site FPR is reported inconsistently (0.03% in Section 6 vs 0.03 with TNR 0.97 in Section 9.2), and the event-level cross-site metrics rest on a single ten-minute recording. · noted in this summary

Read this if you need a one-class acoustic detector for a single target species that needs few labels in open-set tropical soundscapes, or want to see Perch embeddings and an OCSVM linked to occupancy estimates.

Citations
1
FWCI
3.92
Field percentile
top 12%

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISLewis '22Bowler '25Pickering '26Owens '26Fortune '26Gros '26
72% semantic matchEcology
T. C. Lewis, I. Gutierrez Vargas, A. Beckerman+1 more·bioRxiv·Dec 29, 2022