aria-inference-birdnet 0.1.4

BirdNET-only inference package for ARIA (Acoustic Recognition for Inventories of Aviaries)

aria-inference-birdnet

BirdNET-only inference package for ARIA (Acoustic Recognition for Inventories of Aviaries).

This is the baseline detection system for the BioDCASE 2026 Challenge.

Installation

Requires Python 3.11 or 3.12.

pip install aria-inference-birdnet

This automatically installs BirdNET-Analyzer and all other dependencies.

Quick start

1. Download model files

aria-inference-birdnet download-models --dir ./models

This fetches the custom BirdNET classifier and species labels from GitHub Releases.

2. Run detection on an aviary

Each aviary has a different set of species. The aviary_config.json (provided with the challenge data) maps each aviary ID to its species list, so that predictions are filtered to only the species that can actually be present.

aria-inference-birdnet detect \
  --input aviary_1/ \
  --output predictions_aviary_1.csv \
  --model-dir ./models \
  --aviary-config aviary_config.json \
  --aviary aviary_1

3. List available aviaries

aria-inference-birdnet list-aviaries --aviary-config aviary_config.json

4. Python API

from pathlib import Path
from aria_inference_birdnet import run_detection

run_detection(
    input_path=Path("aviary_1/"),
    output_csv=Path("predictions_aviary_1.csv"),
    model_dir=Path("models/"),
    aviary_config=Path("aviary_config.json"),
    aviary_id="aviary_1",
)

How aviary filtering works

The aviary_config.json contains two things:

• Wild birds: species common to all aviaries (e.g. House Sparrow, Common Blackbird). These are always included in predictions.
• Per-aviary species: species that live in a specific aviary.

When you specify --aviary aviary_1, the detector only keeps predictions for that aviary's species plus wild birds. This removes false positives from species that are not present in the aviary.

Alternatively, you can use a flat species file with --allowed-species-file instead of the aviary config. This applies the same species list to all input audio regardless of aviary.

CLI reference

detect

Flag	Default	Description
--input	(required)	Audio file or directory
--output	(required)	Output CSV path
--model-dir	(required)	Directory with model files
--aviary-config	None	Path to aviary_config.json
--aviary	None	Aviary ID (e.g. aviary_1). Requires --aviary-config
--allowed-species-file	None	Flat species whitelist (alternative to aviary config)
--min-confidence	0.05	Global confidence cutoff
--temperature	1.8	Confidence scaling (T>1 reduces saturation)
--overlap	0.0	Window overlap in seconds
--threads	4	CPU threads
--topk	3	Max species per window (0=disable)
--eval-csv	(auto)	Per-class threshold CSV

download-models

Flag	Default	Description
--dir	(required)	Local directory for model files
--skip-optional	False	Skip the evaluation CSV

list-aviaries

Flag	Default	Description
--aviary-config	(required)	Path to aviary_config.json

Pipeline stages

1. BirdNET analysis – runs birdnet_analyzer.analyze on the input audio
2. Species filtering – keeps only species present in the aviary (from config) or in the allowed-species file
3. Temperature scaling – calibrates over-saturated confidences (T=1.8 by default)
4. CSV merging – combines per-file result CSVs into one
5. Per-class thresholds – applies optimal thresholds from training evaluation (if available)
6. Top-K filtering – keeps at most K species per 3-second window

Output format

File name,Start (s),End (s),Scientific name,Common name,Confidence
recording_001.wav,0.0,3.0,Phoenicopterus roseus,Greater Flamingo,0.8734
recording_001.wav,3.0,6.0,Bostrychia hagedash,Hadada Ibis,0.7521

Notes

• This package is inference-only. It does not include training, data fetching, or segmentation.
• Model binaries are hosted as GitHub Release assets to keep the pip package small.
• Python 3.10 is not supported because birdnet-analyzer requires ≥3.11.

License

Apache-2.0