OPEN RESEARCH · MAY 2026

BirdNET Australian Species
Benchmark Report

Field evaluation of BirdNET acoustic AI performance on Australian species. 104 species tested across real-world conditions using Wild Echo Sentinel hardware.

↓ Download Full Report (PDF) View on GitHub

Key Results

What We Found

40.4%

Detection coverage across tested Australian species

42 of 104 species reliably detected

104

Australian species evaluated from eBird AU checklist

Representative sample across habitat types

Datasets — all open access, reproducible

Frozen 2026-05-26, SHA256 verified

v4→v2.4

BirdNET model improvement trajectory documented

Significant gains in recent versions

Analysis

Key Findings

🦜

Coverage Gap

BirdNET excels at species with global training data. Many Australia-endemic species — particularly those in the Maluridae and Meliphagidae families — fall below reliable detection thresholds due to limited global representation in training data.

⚠️

Threatened Species Risk

Several threatened and endangered Australian species showed low detection rates, raising flags for conservation monitoring applications. Detection rates for threatened species are documented separately in the report.

📈

Model Improvement

BirdNET v2.4 shows measurable improvement over earlier versions for Australian species. The improvement trajectory suggests continued gains, but AU-specific fine-tuning remains the most viable path to high coverage.

🔬

Precision vs Coverage

For detected species, precision is high. The primary challenge is coverage — species not reliably detected at all. This points to a training data gap rather than a fundamental model architecture limitation.

Data Visualisation

Results at a Glance

Fig. 1 — Precision vs Coverage: detected species cluster in high-precision band

Fig. 2 — Detection coverage by species family across the 104-species test set

Fig. 3 — BirdNET AU performance improvement trajectory across model versions

Fig. 4 — Detection rates for threatened and endangered Australian species

Open Data

Download the Dataset

benchmark-2026 · Frozen Dataset

✓ SHA256 VERIFIED

All datasets are frozen and publicly available. SHA256 checksums provided for integrity verification. This is v2.1 — the immutable baseline. Future versions will be tagged separately.

📄 benchmark-report-v2.1.pdf PDF 📊 birdnet-au-benchmark-2026-05-14.md Dataset 📊 birdnet-au-benchmark-batch2-2026-05-14.md Dataset 📊 birdnet-au-benchmark-v2-2026-05-14.md Dataset 🔗 github.com/Wild-echo/benchmark-2026 GitHub

Reproducibility

Methodology

Species Selection

104 Australian bird species selected from the eBird AU checklist, weighted toward species with ecological or biosecurity significance. Includes threatened, invasive, and indicator species.

Hardware & Recording

Audio captured using Wild Echo Sentinel hardware in Australian field conditions. Recordings processed at 48kHz. Standardised 60-second segments used for model evaluation.

Model Evaluation

All recordings processed through BirdNET v2.4. Detection threshold set at 0.7 confidence. Species flagged as "detected" only when meeting threshold across multiple recordings.

Dataset Freeze & Verification

Datasets frozen 2026-05-26. SHA256 checksums generated and published. All source data available on GitHub for independent reproduction.

Want to use Wild Echo
on your property?

Wild Echo Sentinel detects 6,000+ species, hive health, and biosecurity threats from a single passive recording. No expertise required.

View Pricing Talk to Us

BirdNET Australian SpeciesBenchmark Report