aria-inference

Full ARIA inference package: BirdNET + PERCH v2 + Fusion for bird species detection in zoo aviaries.

This is the complete baseline system for the BioDCASE 2026 Challenge. For BirdNET-only inference, see aria-inference-birdnet.

Installation

Requires Python 3.11 or 3.12 and a working installation of FFmpeg.

pip install aria-inference

This automatically installs BirdNET (via birdnetlib), PERCH v2 (via perch-hoplite), TensorFlow, and PyTorch.

Note: PERCH v2 auto-downloads its model from Kaggle on first use (~500 MB). This happens once and is cached locally.

Quick start

1. Download model files

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

This fetches from GitHub Releases:

• ZooCustom_v1.tflite – custom BirdNET classifier
• ZooCustom_v1_Labels.txt – species labels
• fusion_model_perchv2_best.pth – trained fusion model
• architecture_info_public.json – model architecture config
• perch_v2_zoo_species_mapping.json – BirdNET-to-PERCH species mapping

2. Run detection on an aviary

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

4. Python API

from pathlib import Path
from aria_inference import run_full_inference

run_full_inference(
    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 it works

ARIA uses a three-model ensemble for robust species detection:

Audio (3-second segments)
  │
  ├─ BirdNET (custom TFLite) → 87-dim probability vector
  │
  ├─ PERCH v2 (foundation model) → 87-dim constrained logits
  │
  └─ Fusion model (PyTorch) → combines BirdNET + PERCH → predictions
      │
      └─ Ensemble voting merges all three → final species + confidence

Two inference modes are available:

• Ensemble voting (default, --mode voting): all three models run on every segment; predictions are merged using weighted voting with a consensus boost when 2+ models agree.
• Tier-based fallback (--mode tiers): BirdNET runs first; if confidence < threshold, fusion runs; if fusion is also low, PERCH provides a fallback.

CLI reference

detect

download-models

list-aviaries

Output format

File name,Start (s),End (s),Species,Confidence,Method
recording_001.wav,0.00,3.00,Greater Flamingo,0.8734,Ensemble Voting
recording_001.wav,3.00,6.00,Hadada Ibis,0.7521,Ensemble Voting

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.
• PERCH v2 auto-downloads from Kaggle on first use and is cached locally.
• GPU is optional but recommended for faster PERCH inference.
• Python 3.10 is not supported due to dependency constraints.

License

Apache-2.0