Bukit Dinding Birdsound Atlas

One recording day · 88 clips · 13 species

From a pre-dawn phantom call to a dusk chorus, one day of sound becomes a living map.

Click or tap a dot to hear a moment from the forest, or start with the dawn chorus.

Audio options

Preview clips on hover

Play a clip to unlock hover previews.

How to read the hero wheel

Each dot marks a 3-second BirdNET detection window, not a confirmed field identification. Playback uses a longer context clip.
The wheel follows your confidence threshold and any species filters you’ve applied.
The wheel always shows the full day, even when other views are focused on a smaller time window.
Click a dot to play its full clip in the listening bar below.
Preview clips are optional and stay off by default. Play any clip once to enable them.
This atlas covers one day at one site. It is not a census.

Soundscape Timeline

Rows are species; dots are clips over time.

Richness (current filters) Uses confidence + species filters; not affected by time brush.

Confidence threshold

These labels are rough guidance; listen to clips to validate.

Higher = fewer, cleaner clips; lower = more, noisier.

≥55%

Showing 69 of 88 clips across 10 of 13 species

Context

The Green Lung

An island of sound in the urban sprawl.

Bukit Dinding constitutes a vital ecological asset within Kuala Lumpur. Encompassing more than 300 acres and rising 291 meters, this former rubber estate has naturally regenerated into a thriving secondary forest. It is early May; the air is 24°C with heavy 90% humidity. To understand who lives in this green lung, we just have to listen. Every dot on this timeline is an bird sound clip flagged by a neural network over a single 24-hour period.

Scroll to follow the day

Detected

The Midnight Phantom

A haunting hoot in the darkness.

Listen for: A soft, rhythmic hoot—"hoo-hoo-hoo"

It is 03:55 AM. The forest is a world of shadows. Then, cutting through the stillness: the hoot of a Collared Scops-owl. The call is low, deliberate, a whisper from the treetops. It lingers, then fades, leaving only the hum of insects and the rustle of unseen movement.

Detected

The First Awakening

The undergrowth stirs before the sun rises.

Listen for: Hard, chopping "chonk-chonk-chonk" notes, repeated in steady bursts.

At 06:50 AM, the true forest residents begin to vocalize. The Pin-striped Tit-Babbler, a small, highly vocal bird of the dense understory, is the first to claim the acoustic space. Moving in agitated little flocks through the vines and bamboo, they cut through the lingering nocturnal hum of insects.

Detected

The Dawn Chorus

The chorus changes shape as the canopy catches the light.

Listen for: A bubbly, clipped chirruping song ("whip wi-wiu").

By 07:05, the first rush has started to loosen. The Pin-striped Tit-Babblers are still working the undergrowth, but now Olive-winged Bulbuls cut in from the forest edge. It feels less like a single outburst than a handoff: the hard chopping notes of dawn giving way to brighter, more fluid calls as the hill begins to warm.

Natural history

The Canopy Watcher

Activity shifts higher into the trees as morning progresses.

Listen for: A dry, harsh "skatch-skatch" delivered in a short, scolding series.

At 08:35, the algorithm isolates the Large Woodshrike. A grey-and-white resident of the mid-storey and canopy, this bird often moves with mixed-species foraging flocks. When disturbed, its harsh warning calls carry clearly down from the treetops to the forest floor.

Detected

The Midday Contraction

The soundscape thins in the equatorial heat.

Listen for: A sharp, isolated, metallic "ink" note.

By 12:20 PM, the temperature peaks at 33°C (91°F). The rules of the forest change. The timeline opens up into longer silences, with only scattered calls breaking through the heat. The dense morning chorus is replaced by the sparse, ringing calls of the Asian Glossy Starling—a sleek, red-eyed urban adapter moving between fruiting trees in the midday glare.

Detected

The Afternoon Break

A brief metallic flare in the long, hot quiet.

Listen for: A clean, metallic "ink" note, followed by another a few seconds later.

By 16:05, the day has gone slack with heat. Then two short notes ring out from an Asian Glossy Starling. It is a small moment, almost easy to miss, but that is what the late afternoon sounds like here now: not a chorus, just quick signals from birds moving between fruiting trees before evening gathers.

Data literacy

Dusk's Second Bloom

A final, quieter surge before night falls.

Listen for: The cheerful, rolling "gloo-gloo-gloo" gurgle of the Yellow-vented Bulbul.

Around 18:00, the oppressive heat lifts and the forest experiences a 'second bloom' of sound. Bulbuls return to vocalize one last time. Notice that the model's confidence here is lower (around 60%). If we raised our strictness threshold to 80%, this dusk chorus would vanish from the chart entirely. In bioacoustics, finding the line between "capturing reality" and "filtering out the noise" is a delicate balance.

Explore

Handing Over the Binoculars

The day belongs to you.

You have followed the pulse of Bukit Dinding through a single rotation of the earth, guided by AI detections. But this is just one path through the data. Now take the controls.

About this Atlas

Method & sources

Core claim

In this one-day Bukit Dinding dataset, bird detections are concentrated in a few prominent periods. Confidence filtering materially changes the atlas by reshaping both the number of visible clips and the mix of species represented.

Definitions

Detection: a BirdNET model hit from a three-second analysis window, linked to a species label, confidence score, and playback clip.
Confidence: the model’s certainty for a detection; it is not the same as ground‑truth verification.
Solar windows: dawn and dusk context windows calculated from site-specific sunrise and sunset times and shown separately from activity peaks.
Species richness (window): the number of unique species detected within a fixed time bin.

Limitations

Data scope & methodology

This atlas visualizes automated detections from a single site on a single recording date; it does not estimate total abundance or occupancy.
Confidence values reflect model certainty, not independent field validation.
Absence of detections should not be interpreted as confirmed absence of a species.

How to read descriptive fields

Vocal descriptions are interpretive: Onomatopoeic renderings (e.g., “whip wi‑wiu”) are human approximations of sound, not exact acoustic transcriptions. Actual vocalizations vary among individuals, regions, seasons, and recording conditions, so descriptions may differ across field guides and audio libraries.
Habitat descriptions are generalized: Habitat and elevational ranges are synthesized from regional accounts of the species. Local populations may use slightly different microhabitats depending on vegetation structure, disturbance history, or landscape context. Land‑use changes may also alter habitat associations over time.
Elevational limits are not absolute: Statements such as “mostly below 500 m” reflect typical occurrence patterns, not strict ecological boundaries. Occasional records outside the stated range are possible.
Site-level likelihood is inferred: Fields such as expected_in_bukit_dinding reflect habitat-based inference, not systematic site survey data. They estimate likely suitability from habitat preference and regional distribution, not standardized abundance sampling at Bukit Dinding.
State-level distribution is checklist-based: State and national distribution summaries are derived from compiled checklists and public databases. These may vary slightly between sources and may be updated as new records are documented. Absence from a list does not necessarily confirm true absence.
Similar-species comparisons are simplified: Distinguishing notes are provided as practical field guidance. Vocal and behavioral traits can overlap, and identification should not rely solely on a single characteristic.

Temporal dynamics

Conditions can change over time: Species distributions, habitat use, and abundance may shift because of urbanization, deforestation, climate variation, or improved survey coverage. Dataset summaries reflect information available at the time of compilation.

Sources & citation

Latest source accessed: Apr 7, 2026

Kahl et al. (2021): BirdNET: A deep learning solution for avian diversity monitoring (opens in a new tab)

Accessed Feb 19, 2026

Primary methodological citation for the BirdNET model family used via BirdNET-Analyzer.
Time and Date: sunrise and sunset in Kuala Lumpur (opens in a new tab)

Accessed Feb 19, 2026

Reference context for dawn/dusk bounds on the recording date.

Species-specific references and image attributions appear inside each Species Spotlight.

How to cite

Bukit Dinding Birdsound Atlas. Slicing / Dicing. https://dataviz-767.pages.dev/2026/02/19/bukit-dinding-birdsound-atlas. Published 2026-02-19. As of 2026-04-18.

Takeaway: A single day of listening at Bukit Dinding is not a constant chorus but a shifting pattern of peaks, lulls, and uncertain signals, and the atlas lets readers explore those changes directly.

Why it matters: Bioacoustic data is rich but often hard to interpret, so this atlas turns timing, species mix, and uncertainty into a format that works for both curious readers and technical users.

Canonical link: https://dataviz-767.pages.dev/2026/02/19/bukit-dinding-birdsound-atlas

Corrections

or review the corrections policy .

Next step

From Raw Data to Ecological Storytelling

This interactive atlas emerged from a bioacoustics workshop co-hosted by Friends of Bukit Dinding (FoBD) and the Sunway Centre for Planetary Health on February 7, 2026.

Passive acoustic monitoring generates large datasets that are often difficult for non-specialists to interpret. By turning BirdNET detections into an interactive atlas of context audio clips and spectrograms, this project helps readers explore patterns in biodiversity within Bukit Dinding, one of Kuala Lumpur’s urban green lungs.

Detection Uncertainty & Scientific Transparency

These detections are probabilistic matches, produced by BirdNET's neural classifier, not independently confirmed by expert human review. In ecological data analysis, there is a trade-off between precision (fewer false positives) and recall (fewer false negatives).

This interface exposes classifier uncertainty directly through the confidence threshold. The confidence slider’s default threshold is 0.55, and you can adjust it as low as 0.1.

Lowering the threshold casts a wider acoustic net: it increases recall and usually surfaces more detections, but with lower reliability. Raising it favors precision and isolates the most convincing matches. The absence of a detection does not mean a species was absent; it may also reflect distance, ambient noise, overlap with other sounds, or gaps in recording coverage.

After the initial BirdNET-based compilation, the published clip set was manually reviewed with the help of an experienced birder to identify likely misclassifications. As a result, the atlas was refined from 102 clips across 21 species to 88 clips across 13 species.

One important takeaway from this process is that BirdNET detections are a useful starting point, not an absolute ground truth. In practice, reviewing clips against reference vocalizations in a noisy urban soundscape proved both valuable and difficult. This published version should therefore be understood as a BirdNET-assisted, human-reviewed exploratory dataset rather than a direct dump of raw classifier output.

Data Processing Pipeline

To ensure traceability from the field to the browser, the data architecture relies on a deterministic Python build script, build_birdsound_assets.py. All parsed detections are included in the dataset; the script does not discard detections below the default threshold.

Field recordings and detection: Long-duration passive acoustic WAV files (for example, 2MM04211-MT1_20250503_000000.wav) are analyzed with BirdNET-Analyzer. The build pipeline ingests all detections and preserves the full range of confidence scores.
Audio extraction: BirdNET detection windows remain 3 seconds, while playback clips are exported as longer listening context where available (in this build, most are 8 seconds, with shorter clips near source boundaries). Clips are converted to MP3 using ffmpeg in mono at 48 kHz and 160 kbps, with natural volume dynamics preserved (normalization: none).
Feature extraction and visuals: The pipeline generates WebP spectrograms via matplotlib and librosa. It computes key acoustic features including dominant_frequency_hz, bandwidth_hz, and log-energy frequency profiles (500–12000 Hz) to derive the PCA embedding coordinates used in the "Sound Constellation" view.
Reproducibility: Every data point carries a deterministic ID mapped back to the original day timeline, supporting full traceability.

Bukit Dinding Birdsound Atlas

The Green Lung

The Midnight Phantom

The First Awakening

The Dawn Chorus

The Canopy Watcher

The Midday Contraction

The Afternoon Break

Dusk's Second Bloom

Handing Over the Binoculars

About this Atlas

Method & sources

Data scope & methodology

How to read descriptive fields

Temporal dynamics

How to cite

Corrections

Next step

From Raw Data to Ecological Storytelling

Detection Uncertainty & Scientific Transparency

Data Processing Pipeline

Acknowledgements

Support