Dolphin broadband transient vocalization detection in impulsive ambient noise

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

cs.SDOct 26, 2026

Dolphin broadband transient vocalization detection in impulsive ambient noise

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

Organizations: Acoustic Research Laboratory, Tropical Marine Science Institute, National University of Singapore. · Department of Electrical and Computer Engineering, National University of Singapore.

Abstract

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.

Figures & tables

Fig. 1 : Representative timeseries and corresponding spectrograms of high-SNR examples of the two target signal classes considered in this work, recorded in controlled conditions. (a) and (c) Clicks are characterized by isolated broadband impulsive peaks arranged in click trains, while (b) and (d) burst-pulses exhibit more extended structure with denser temporal organization.

Fig. 2 : (a) Representative timeseries and (b) spectrogram of snapping shrimp noise background showing broadband impulsive nature of the snaps.

Fig. 3 : Map of island sites in Singapore’s southern waters where passive acoustic recorders were deployed for the study, with circles whose size is proportional to number of days with detections at these sites normalized to total number of deployment days at the site, and (inset) zoomed out map of the region around Singapore showing the area studied marked by a red box.

Location	No. of	No. of	Months
	broadband signals	days with	of data
	detected	detections	recorded
Hantu	170	33	7.7
Jong	13	7	5.2
Kusu	14	3	4.3
Raffles Lighthouse	61	5	5.2

TABLE I : Island-sites at which acoustic data was recorded, with number of broadband transient vocalizations detected and number of days on which these were detected.

Fig. 4 : Architecture of the proposed 1D CNN detector TRIDENT. The network operates directly on a 500 ms raw waveform inputs and uses a deep but narrow design with one residual block, progressive temporal downsampling, adaptive average pooling, and a lightweight classification head. Symbols: $k$ is the kernel size, $s$ is the stride length, $ch$ is the number of filters, B is batch-size

Split	Pos	Pos	Neg	Neg	Total
	(Click)	(Buzz+Burst-pulse)	(Noise)	(Hard)	independent
					samples
Train	723	808	3390	93	5014
Validation	187	196	700	33	1116
Test	491	247	1893	33	2664

TABLE II : Composition of the final mixed-target dataset used in the mainline configuration of this work. Pos indicates independent click, buzz and burst-pulse samples, which are mixed with noise in different permutations to produce samples for training and validation, and Neg indicates independent negative data samples (non-signals).

Fig. 5 : Comparison of RMS distributions of positive and negative sample windows under normalization based on (a) noise background alone (for positives), and (b) proposed normalization based on data. Under the proposed strategy, the overlap between the two distributions increases substantially, encouraging the detector to rely on acoustic structure rather than global amplitude shortcuts.

Fig. 6 : ROC comparison of the click-only, buzz+burst-pulse-only, and joint models on the final mixed-test set. The annotated points indicate PD@PFA=1% for each model. The joint model achieves the best overall mixed-target performance.

Fig. 7 : Two examples of hard negatives introduced into the training dataset with features similar to buzzes or squawks, (a) and (b) spectrograms of hard negative clips containing fast boat sounds, and (c) and (d) their corresponding timeseries waveforms showing regular broadband impulsive structure.

Model / Data for metric	Click-only PD@PFA=1%	Buzz+burst-pulse PD@PFA=1%	Mixed-test PD@PFA=1%	Mixed-test AUC
Click-only	0.786	0.514	0.695	0.936
Buzz+burst-pulse	0.092	0.903	0.363	0.854
Click + buzz + burst-pulse	0.811	0.911	0.844	0.959
Click + buzz + burst-pulse (no hard negatives)	0.752	0.870	0.791	0.974
Surfperch + Logistic regression	-	-	0.753	0.966
Surfperch + MLP	-	-	0.782	0.969

TABLE III : Comparison of different ML-based detectors and target definitions under click-only, buzz+burst-pulse-only, and final mixed-test settings. The mixed-test results are reported on the final evaluation set consisting of all three signal types, plus the hard negatives. Bold values in a column indicate the largest row in that column

Underwater Acoustics

See all 23 papers

Ultra-Compact CNN Architectures for Tropical Bird Audio Detection on Microcontrollers
Jul 22, 2026Muhammad Mun'im Ahmad Zabidi, Mohd Yamani Idna Idris, Norisma IdrisConvolutional Neural Networks Wildlife Monitoring
Faculty of Computer Science and Information Technology, Universiti Malaya, Kuala Lumpur, Malaysia
OpenWhistle: A Large-Scale Longitudinal Dataset and Benchmark of Bottlenose Dolphin Vocalizations
Sep 28, 2026Faadil Mustun, Chiara Semenzin, Roberto Dessi +7Sound Event Detection Bioacoustics
Institut de Biologie de l’École normale supérieure, CNRS, INSERM, Université PSL, Paris, France · Earth Species Project, France · Not Diamond, San Francisco, USA +4
Time-frequency localization of bird calls in dense soundscapes
Jun 9, 2026Simen Hexeberg, Fanghui Tong, Hari Vishnu +1Sound Event Detection Object Detection
Tropical Marine Science Institute, National University of Singapore, Singapore · Dept. of Electrical and Computer Engineering, National University of Singapore, Singapore · School of Computing, National University of Singapore, Singapore

Jul 22, 2026J/K move · Enter open · S save

Ultra-Compact CNN Architectures for Tropical Bird Audio Detection on Microcontrollers

Muhammad Mun'im Ahmad Zabidi, Mohd Yamani Idna Idris, Norisma Idris

Faculty of Computer Science and Information Technology, Universiti Malaya, Kuala Lumpur, Malaysia

Convolutional Neural Networks Wildlife Monitoring Sound Event Detection Tiny ML Bioacoustics Passive Acoustic Monitoring+1

Open paper AskPDF arXiv

Passive acoustic monitoring of tropical biodiversity is bottlenecked by the storage and battery cost of continuously recording soundscapes in which bird vocalisations typically occupy less than 10% of the audio. Autonomous recording units built on low-power microcontrollers (typically ARM Cortex-M with

\leq

256 kB of RAM) address this by triggering only on likely-positive segments, but the on-device options are unsatisfying: coarse frequency-energy triggers such as Goertzel filters flood SD cards with false positives at

\sim

71% precision, whereas neural detectors developed for temperate single-species tasks are either too large to deploy or transfer poorly to species-rich tropical settings. We present DrongoNet, a family of three INT8 CNN detectors sized for this envelope and validated on a 50,000-clip, 1,677-species Southeast Asian tropical dataset (SEABAD). The headline model, DrongoNet-Micro (919 parameters, 6.26 kB, 0.9810 AUC, 98.3% mean recall at τ = 0.35), is a drop-in replacement for the Goertzel trigger used in commodity field recorders: at α = 0.10 tropical prevalence it captures 8 pp more bird vocalisations than Goertzel and extends a 32 GB card from

\sim

28 to

\sim

45 days of monitoring. DrongoNet-Nano (5.09 kB) bounds the ultra-low-flash extreme; DrongoNet-Edge (33.06 kB, 0.9991 AUC) targets Linux SBCs. On SEABAD, Micro matches a retrained TinyChirp CNN-Mel baseline within 0.1 pp AUC at 28

\times

fewer parameters, confirming that the family is deployment-agnostic across mel-spectrogram bird corpora but requires per-environment retraining. Full INT8 quantisation costs

<

0.12% AUC across all three variants.

Abstract

Underwater Acoustics

Explore similar work

Ultra-Compact CNN Architectures for Tropical Bird Audio Detection on Microcontrollers

OpenWhistle: A Large-Scale Longitudinal Dataset and Benchmark of Bottlenose Dolphin Vocalizations

Time-frequency localization of bird calls in dense soundscapes