PaperlayerView exact revision on arXiv

Source-linked AI summary

Dolphin broadband transient vocalization detection in impulsive ambient noise

Hari Vishnu, Jin Xinhui, Mandar Chitre, Simen Hexeberg, Koay Teong Beng

arXiv:2610.35641v1cs.SDeess.ASeess.SP

TL;DR

Dolphin clicks, buzzes and burst-pulses are difficult to detect in tropical waters because they share broadband impulsive structure with snapping-shrimp noise. The paper develops TRIDENT, a waveform-based detector, and reports detections across long-term Singapore recordings, with potential for real-time edge deployment.

  • Problem

    Broadband dolphin transients overlap with snapping-shrimp noise, while reliable local population and visitation estimates remain lacking.

    Source: §I · p. 1

  • Method

    TRIDENT uses a raw-waveform 1D CNN with controlled SNR-based mixing, waveform-specific normalization and hard negatives.

    Source: §I · p. 2, §IV · p. 4, E. Data sources · p. 6

  • Results

    325 vocalization events in 0.5 s windows were detected from June 2019 to September 2021, with relatively heavier activity at Hantu, Sisters Islands and Raffles Lighthouse.

    Source: §VI · p. 10

  • Takeaways & Limitations

    The lightweight detector’s inference-time performance and 1% false-alarm load are described as compatible with quality-control review and potential real-time edge-based monitoring.

    Source: C. Practical implications and future… · p. 10

Abstract

from arXiv · show

Passive acoustic monitoring in tropical waters such as those in Southeast Asia is hampered by the persistent, high-level impulsive noise of snapping shrimp. This is a particularly daunting challenge for detecting broadband transient vocalizations of dolphins - clicks, buzzes and burst-pulses. These vocalizations, unlike frequency-modulated whistles, share the broadband, impulsive character of the shrimp snaps and cannot be easily separated from them by frequency-localized or energy-based means. We develop TRIDENT, a lightweight waveform-based detector for these vocalizations, and apply it to 54.2 recorder-months of data from ten island sites around Singapore. We show that the training-data pipeline and the architecture are both important: controlled signal-to-noise-ratio based mixing to generate training data, appropriate normalization and hard negative mining, together steer the detector toward learning transient temporal structure rather than energy shortcuts. A single joint detector handles all three signal types better than specialized models for each, and outperforms established baselines. Applied to long-term recordings, it reveals tightly clustered, predominantly daytime activity, motivating days-with-detections as a robust indicator of cetacean presence.

I. INTRODUCTION

Passive acoustic monitoring can help study dolphin populations, but broadband transient vocalizations remain difficult to detect in shrimp-dominated waters. TRIDENT addresses this challenge with a waveform-based detector and training strategy designed to preserve and learn transient structure.

I. INTRODUCTION

  • Singapore lacks reliable estimates of local dolphin populations or visitation patterns, while such information is important for conservation and policy-making.Passive acoustic monitoring offers wide-area, around-the-clock coverage, but its large recording volumes make automated detection essential.Source: §I · p. 1
  • Snapping shrimp produce high-level broadband impulsive noise that overlaps with dolphin clicks, buzzes and burst-pulses, making frequency-localized or energy-based separation difficult.These vocalizations share short duration, broadband content and impulsive structure with the dominant interference.Source: §I · p. 1
  • TRIDENT is a waveform-based 1D CNN for click trains, buzzes and burst-pulses, applied to 54.2 recorder-months from ten Singapore island sites.The approach avoids the nonlinear denoising used for whistles and emphasizes a data-centric training strategy alongside the architecture.Source: §I · p. 2

II. CHALLENGES AND EXISTING METHODS

  • Detection is challenged by scarce high-quality positives, low-SNR deployment recordings and shrimp snaps that resemble the target signals.The task is also sensitive to normalization, clipping and mixing: poor choices can remove useful waveform information or create an artificial energy gap that invites shortcuts.Source: §II · p. 2

A. Prior work

Earlier click-detection approaches use signal-processing features or spectrogram-based learning, but impulsive shrimp noise remains a difficult confound. TRIDENT addresses the gap with supervised waveform-domain learning and a data-centric training framework.

A. Prior work

  • Heuristic energy and entropy methods can degrade under snapping-shrimp interference, while predefined time-frequency features may limit capture of complex or unseen patterns.Spectrogram-based CNNs and other deep-learning approaches have been applied to marine mammal detection, but the cited examples do not address this specific transient-detection gap.Source: A. Prior work · p. 2
  • Broadband transient detection in shrimp-dominated shallow tropical waters remains unaddressed by the prior Singapore detector, which targets frequency-modulated signals.Energy-, feature- and frequency-localized approaches are insufficient when shrimp noise and dolphin transients overlap strongly.Source: A. Prior work · p. 3
  • TRIDENT combines supervised waveform-domain learning with controlled SNR-based data generation, waveform-specific normalization, hard negatives and training-time augmentation.The framework is intended to address limited training data and reduce reliance on artificial statistical cues that may not generalize to deployment recordings.Source: A. Prior work · p. 3

III. EXPERIMENTAL SETUP AND DATA COLLECTION

The study collected passive acoustic recordings at ten island sites in Singapore’s southern waters over several months between June 2019 and September 2021. The dataset totals 54.2 recorder-months.

III. EXPERIMENTAL SETUP AND DATA COLLECTION

  • The study’s map locates the recorder sites, and its table reports detected vocalization counts and detection-day counts for the island sites.The map’s circle sizes encode detection days normalized by each site’s deployment days.Source: Figure 3 · p. 4, Table I · p. 4
  • Recorders were deployed intermittently at ten island sites between June 2019 and September 2021, yielding 54.2 recorder-months of data.Each recorder was a bottom-deployed stand-alone unit, typically deployed for approximately one month per deployment.Source: §III · pp. 3–4

IV. DETECTOR DESIGN AND TRAINING

TRIDENT operates directly on raw waveforms to preserve the transient structure needed to distinguish dolphin vocalizations from impulsive background noise. Its design combines preprocessing choices, a lightweight 1D CNN and a 500 ms input window.

IV. DETECTOR DESIGN AND TRAINING

  • The detector design coordinates raw-waveform 1D CNN classification, controlled SNR-based mixing, waveform-specific normalization and hard negatives.These choices aim to preserve transient peaks while avoiding artificial class separation and improving robustness to realistic false positives.Source: §IV · p. 4

A. Preprocessing

  • Raw audio is resampled to 44.1 kHz and high-pass filtered at 1 kHz to suppress low-frequency background and emphasize high-frequency transient content.The sampling rate reflects the availability of training data at that rate and the observed concentration of detectable open-ocean signal energy up to 30 kHz.Source: A. Preprocessing · p. 4
  • The pipeline avoids nonlinear pre-filtering because denoising can flatten the transient peaks that distinguish clicks, buzzes and burst-pulses.This preserves waveform fidelity while leaving shrimp snaps present in the input and shifting discrimination to the model.Source: A. Preprocessing · p. 4

B. Detector architecture

  • A lightweight 1D CNN processes fixed-length 500 ms raw-waveform segments, or 22,050 samples, to retain transient patterns and their temporal organization.The window balances enough context for repeated-pulse structure against computational cost and the risk of including unrelated events.Source: B. Detector architecture · pp. 4–5

C. Binary detection classes

The task classifies waveform segments as dolphin broadband vocalizations or non-target sounds, with one joint detector for clicks, buzzes, and burst-pulses performing on par with specialized detectors. The model uses a raw-waveform CNN, and evaluation emphasizes detection at a low false-alarm rate.

C. Binary detection classes

  • The binary classifier distinguishes target segments containing dolphin broadband vocalizations from non-target background noise and other sounds.Its output is a detector score that can be thresholded into a detection decision.Source: C. Binary detection classes · p. 5
  • TRIDENT processes fixed-length 500 ms raw waveforms with a deep, narrow CNN, one residual block, temporal downsampling, adaptive average pooling, and a lightweight head.The architecture is described in Fig. 4.Source: Figure 4 · p. 5
  • The joint formulation performs on par with specialized detectors while requiring only one model at inference.The joint model detects clicks, buzzes, and burst-pulses in one binary framework.Source: C. Binary detection classes · p. 5

D. Model training and evaluation

  • Evaluation reports recall at a 1% probability of false alarm and ROC AUC, emphasizing the low-false-alarm operating regime for long-term monitoring.A low false-alarm rate helps keep shortlisted detections manageable for downstream review and analysis.Source: D. Model training and evaluation · p. 6

E. Data sources

The dataset combines cetacean vocalization recordings from varied species and acquisition conditions with real underwater noise and challenging non-target examples. Training and validation positives use synthetic mixing, while the test positives are mostly low-SNR clips from the original recordings.

E. Data sources

  • The positive sources span varied species, environments, and acquisition conditions, with emphasis on Southeast Asian species; high-SNR clips mostly support training and validation, while low-SNR clips shape the test set.Sources include recordings and open repositories covering multiple dolphin species and porpoises.Source: E. Data sources · p. 6
  • Background negatives come from real underwater recordings, typically containing shrimp noise and sometimes ships.The final mixed-target dataset combines independent positive samples with noise permutations and independent negative samples.Source: E. Data sources · p. 6, Table II · p. 6
  • Hard negatives are false positives mined and manually verified, typically non-target signals with buzz-like, impulsive, or high-frequency transient characteristics.The final experimental dataset used 159 hard-negative segments.Source: E. Data sources · p. 6
  • After initial detections, annotators reviewed the highest-scored 100 clips per day, labeled true positives and false alarms, and added annotations to the training/validation dataset.They also listened near true positives for possible missed detections.Source: E. Data sources · p. 6
  • The split was roughly 50:15:35 for training, validation, and test, with adjustments to include each positive subclass; validation positives were synthetically mixed, unlike test samples.The validation set has fewer raw samples than the test set because of this additional mixing.Source: E. Data sources · p. 6

F. Training data preparation and augmentation

Training and validation windows are generated by repeatedly mixing available vocalization recordings with noise at controlled SNRs. The signal term uses peak amplitude to reflect transient structure, and the SNR range is jointly chosen with that definition.

F. Training data preparation and augmentation

  • Synthetic data are generated by randomly mixing available positive signals with noise in each iteration using controlled SNR-based mixing.This approach addresses the scarcity of recordings typical of Singapore waters and increases data diversity.Source: F. Training data preparation and… · pp. 6–7
  • About 3.6 million unique signal-plus-noise permutations support training and validation, with shifting and amplitude scaling used as further augmentation.Flipping is omitted because the target vocalizations and shrimp snaps are not time-symmetric, and temporal information may aid detection.Source: F. Training data preparation and… · p. 7
  • The signal SNR term uses the peak of the peak-normalized positive waveform against the noise segment's RMS, reflecting transient peaks rather than whole-window signal energy.The peak-based signal measure is chosen because discriminative information is concentrated in short-duration, high-amplitude events.Source: F. Training data preparation and… · p. 7
  • Clicks use peak-to-noise SNRs of [15.6, 35.6] dB, while burst-pulses and buzzes use [20.4, 40.4] dB.The authors state that the SNR definition and its range must be designed jointly; these ranges correspond to smaller values under typical RMS-based calculations.Source: F. Training data preparation and… · p. 7

G. Normalization

Normalization is designed to reduce energy-based shortcuts while preserving waveform dynamics. Training-only amplitude scaling adds magnitude variation, and held-out validation and test sets serve distinct purposes.

G. Normalization

  • Positive mixed segments and negative segments are normalized by median absolute deviation, keeping their RMS distributions substantially overlapping so energy alone cannot separate the classes.The approach avoids clipping and nonlinear denoising, preserving waveform dynamics end-to-end.Source: G. Normalization · p. 7
  • Training waveforms receive dynamic random amplitude scaling from [0.8, 1.2] without changing SNR; validation and test data remain unchanged.The augmentation is intended to encourage scale-robust features rather than reliance on absolute signal magnitude.Source: G. Normalization · p. 7
  • The held-out test set contains 2,900 samples, with mostly low-SNR positives selected from original recordings; the 5,980-sample validation set uses synthetically mixed positives to guide stopping.Training stops when validation recall consistently decreases or saturates.Source: G. Normalization · p. 7

V. RESULTS

  • The results section first checks for an exploitable energy shortcut, then evaluates joint versus signal-specific definitions, hard negatives, and comparisons with established methods.This passage outlines the order of the results analyses.Source: §V · p. 7

A. Effect of normalization

The proposed normalization closes an energy gap between positive and negative samples, steering detection toward waveform structure. In mixed-target evaluation, the joint click+buzz+burst-pulse detector outperforms specialized models and is adopted for TRIDENT.

A. Effect of normalization

  • The proposed normalization substantially increases positive-negative RMS overlap, preventing separation by global amplitude alone.Background-only normalization leaves a pronounced energy gap; the proposed strategy normalizes after mixing.Source: A. Effect of normalization · pp. 7–8, Figure 5 · p. 8

B. Impact of target definition and detector design

C. Impact of hard negative samples

Hard negatives improve detection, while TRIDENT is benchmarked against signal-processing and Surfperch-based methods. It exceeds the stronger Surfperch models at the low-false-alarm operating point and runs quickly on an edge-computing module.

C. Impact of hard negative samples

  • Hard-negative training yields about 4% higher PD than training without hard negatives.These challenging non-targets include boat- and ship-generated sounds with broadband impulsive structure resembling dolphin vocalizations.Source: C. Impact of hard negative samples · p. 9

D. Real-time performance and comparison with existing methods

  • Inference takes an average of 8.4 ms per 0.5 s audio window, including preprocessing and model inference, on a Jetson Xavier NX.The benchmark tested five minutes of audio.Source: D. Real-time performance and… · p. 9
  • Surfperch-based comparison models use frozen embeddings with logistic-regression or MLP classifiers and contain 24.2–24.5 million parameters.They are approximately 4500 times larger than TRIDENT, which has 5362 parameters.Source: D. Real-time performance and… · p. 9
  • The TKEO detector achieves AUC of 0.30 (PD@PFA=1% of 0.0005), while the entropy-based detector achieves AUC of 0.56 (PD@PFA=1% of 0.0015).Both baselines key on impulsive energy, which is overwhelmingly shrimp-generated in Singapore waters.Source: D. Real-time performance and… · pp. 9–10
  • TRIDENT reaches PD@PFA=1% of 0.844, versus 0.753 for Surfperch+logistic-regression and 0.782 for Surfperch+MLP, with roughly three orders of magnitude fewer parameters.Its AUC is only marginally below the Surfperch-based models.Source: D. Real-time performance and… · p. 10

VI. DISCUSSION AND CONCLUSION

Across the island sites, detections were clustered and predominantly daytime, rather than evenly distributed. The authors therefore recommend days with detections as a more robust indicator of visitation and note that fuller temporal and behavioral analysis remains future work.

VI. DISCUSSION AND CONCLUSION

  • 325 vocalization events were detected across the long-term recordings, with relatively heavier activity at Hantu, Sisters Islands, and Raffles Lighthouse.No vocalizations had yet been detected at Semakau or TPT; the identified hotspots fall within island groups where overlapping dolphin-occurrence hotspots were independently reported.Source: §VI · p. 10

A. Temporal structure of detections

  • Because detections cluster within minutes and are separated by long silent stretches, raw signal counts can be a fragile measure of animal presence.The authors report days with detections as a more robust visitation indicator for comparable long-term PAM studies in this regime.Source: A. Temporal structure of detections · p. 10
  • Almost all broadband vocalizations occurred between approximately 07:00 and 20:00, a daytime pattern consistent with active foraging or social behavior.The authors say dedicated study is warranted; fuller spatio-temporal, cyclical, and behavioral analysis is beyond this methods-focused paper’s scope.Source: A. Temporal structure of detections · p. 10

B. Generalization across species and sites

TRIDENT’s conclusion reports strong detection performance and clustered, predominantly daytime activity, while suggesting days with detections as a robust presence metric. The discussion notes that species generalization remains untested and that future work could address data and model limitations and real-time deployment.

B. Generalization across species and sites

  • TRIDENT is expected to recover IPHD, IPBD, and Irrawaddy dolphin transients most accurately and robustly, but broader species generalization remains untested.The positive set spans many dolphin and porpoise species, but its broader capability has not yet been tested.Source: B. Generalization across species and… · p. 10

C. Practical implications and future work

  • The lightweight raw-waveform 1D CNN is described as well suited to on-device inference, with real-time edge-based PAM proposed as a future direction.At PFA = 1%, its review load is described as consistent with feasible quality-control review and compatible with such deployments.Source: C. Practical implications and future… · p. 10
  • Rule-based candidate selection may bias positives toward prominent or canonical patterns, and low-SNR annotated training data are lacking.The lightweight architecture may also limit capture of complex temporal dependencies; species generalization beyond IPHD and IPBD remains to be tested.Source: C. Practical implications and future… · p. 10
  • The study says its findings can plug gaps in understanding reported delphinid occurrences and behavioral patterns in Singapore’s hyper-urbanized waters.The outputs also provide a basis for forthcoming analyses of spatio-temporal, cyclical, and behavioral patterns and applied downstream tasks.Source: C. Practical implications and future… · p. 11

D. Conclusion

  • A single joint detector handles all three signal types with negligible loss relative to specialized models and substantially outperforms generic signal-processing baselines.Despite being roughly three orders of magnitude smaller, it achieves the best detection probability at a 1% false-alarm rate among the compared methods.Source: D. Conclusion · p. 11
  • Applied across 54.2 recorder-months, TRIDENT revealed tightly clustered, predominantly daytime activity, motivating days with detections as a robust presence metric.The conclusion also reports that controlled SNR-based mixing, normalization, hard negative mining, and amplitude-scaling augmentation steer learning toward transient temporal structure rather than energy shortcuts.Source: D. Conclusion · p. 11
Loading 2610.35641v1…