BSG Finnish Birds Model – ONNX Optimized

Optimized ONNX conversions of the BSG – Finnish Birds Model, a pretrained bird sound classification model fine-tuned for Finland. The original model was developed at the University of Jyväskylä and is based on the BirdNET model architecture.

These are fused models that combine BirdNET's EfficientNet-B0 feature extractor with the BSG classifier head into a single ONNX graph, making them directly compatible with standard BirdNET inference pipelines — no separate feature extraction step required.

Model Description

The BSG Finnish Birds Model uses an EfficientNet-B0 backbone (from BirdNET-Analyzer) with a custom classification head trained on vocalizations of 263 Finnish bird species, covering all breeders, non-breeding migrants, and most common vagrants.

The original BSG model is distributed as a standalone classifier that expects pre-extracted BirdNET embeddings as input. The fused ONNX models in this repository merge BirdNET's feature extractor and the BSG classifier into a single end-to-end model that accepts raw spectrograms and outputs species predictions — identical to how the standard BirdNET ONNX model operates. This makes the BSG model a drop-in replacement for BirdNET in any application that supports custom ONNX models.

Key characteristics:

• Processes audio in 3-second overlapping segments (spectrogram-based)
• Outputs species-wise detection probabilities, calibrated per species via logistic regression
• Predictions are filtered by species' seasonal and geographic plausibility to reduce misclassifications
• Compatible with standard BirdNET inference workflows (same input/output interface)

Files

File	Description	Recommended Use
BSG_birds_Finland_v4_4_fused_fp32.onnx	Full precision (FP32)	GPU (CUDA/TensorRT), Desktop CPU
BSG_birds_Finland_v4_4_labels_fi.txt	Labels (Scientific name_Finnish name)	Class index mapping

ONNX Optimization

The ONNX models were converted from the original TFLite model and optimized for efficient inference using birdnet-onnx-converter. ONNX enables deployment across a wide range of runtimes and hardware platforms including CPU, GPU, and edge devices via ONNX Runtime.

Training Data

Training data combined:

• Expert-annotated clips from Xeno-canto
• Finnish soundscapes
• Targeted field recordings
• Selected mobile phone recordings

Usage

The code for running inference with the model is available at: luomus/bird-identification

For building locally fine-tuned classifiers, see the BSG classifier builder: plauha/BSG_classifier_builder

License

Usage is limited to non-commercial purposes. See the LICENSE file for details.

Citation

If you use this model, please cite:

@article{nokelainen2024mobile,
  title={A Mobile Application–Based Citizen Science Product to Compile Bird Observations},
  author={Nokelainen, O. and Lauha, P. and Andrejeff, S. and H{\"a}nninen, J. and Inkinen, J. and Kallio, A. and Lehto, H.J. and Mutanen, M. and Paavola, R. and Schiestl-Aalto, P. and Somervuo, P. and Sundell, J. and Talaskivi, J. and Vallinm{\"a}ki, M. and Vancraeyenest, A. and Lehti{\"o}, A. and Ovaskainen, O.},
  journal={Citizen Science: Theory and Practice},
  volume={9},
  number={1},
  pages={24},
  year={2024},
  doi={10.5334/cstp.710}
}

@article{lauha2025bsg,
  title={Bird Sounds Global - model builder: An end-to-end workflow for building locally fine-tuned bird classifiers},
  author={Lauha, Patrik and Rannisto, Meeri and Somervuo, Panu and others},
  journal={Authorea},
  year={2025},
  doi={10.22541/au.176599468.88578746/v1}
}

Acknowledgments

This model was developed at the University of Jyväskylä. The ONNX conversion and optimization were performed independently to enable broader deployment options.