mcht67/Polyphonic-BirdSet-train

Polyphonic Bird Call Dataset

Dataset Summary

The Polyphonic Bird Call Dataset is a corpus of synthetic soundscape mixtures generated from individual bird call recordings sourced from BirdSet. Individual recordings are randomly mixed to produce a large collection of polyphonic soundscapes covering varying degrees of polyphony and signal-to-noise ratios. This enables controlled model training and systematic evaluation for tasks such as polyphony estimation, source separation, denoising, and multi-label classification.

The dataset is organized by geographic subset, mirroring the subset structure of BirdSet (e.g. HSN, PER, ...). Additional subsets will be added over time.

Dataset Structure

Data Instances

Each instance in the dataset represents a synthetic audio mixture along with associated metadata describing its composition. Additionally it contains the same mixture without background noise and the source files of the mix.

Data Fields

mixtures:

• audio: Audio mixture file
• no_noise_audio: Audio mixture file without noise
• sources_audio: Audio segments used as sources for mixture
• noise_audio: Noise audio segment
• polyphony_degree: Number of simultaneously vocalizing birds
• birdset_code_multilabel: List of species present in the mixture
• snr_dB: Signal-to-noise ratio of the mixture in dB
• raw_files_*: Metadata of raw files

noise:

• audio: Background noise recording

Subsets

Subset	Region	Status
HSN	(add region description)	Available
PER	(add region description)	Coming soon
(more)	(...)	Coming soon

Subset identifiers follow the BirdSet naming convention. Refer to the BirdSet dataset card for detailed descriptions of each geographic region.

---

Dataset configs

Each subset (e.g. HSN) contains the following configs:

Config	Description
{subset}	Synthetic mixtures of single species segments
{subset}_balanced	Synthetic mixtures from species balanced data (optional)
{subset}_segments	Single species segments used during mixing
{subset}_noise	Background noise recordings used during mixing
{subset}_test	Soundscape recordings for testing

---

Dataset Creation

Source Data

Individual bird call recordings are sourced from BirdSet, which itself aggregates recordings from Xeno-canto and other sources. Please refer to the BirdSet dataset card for detailed licensing and provenance of source recordings.

Generation Pipeline

Mixtures are generated using the pipeline available at: https://github.com/mcht67/Polyphonic-Bird-Call-Dataset

The pipeline randomly selects and mixes individual recordings with controlled polyphony levels and SNR values. Each dataset version is fully reproducible by checking out the exact pipeline commit recorded in the subset metadata.

Reproducing a Specific Dataset Version

Every subset contains a metadata.json file that records the exact pipeline state used to generate it:

{
  "git_commit": "a3f2c1d...",
  "pipeline_repo": "https://github.com/mcht67/Polyphonic-Bird-Call-Dataset",
  "parameters": {
    "random": { "random_seed": 42 },
    "dataset": { "subset": "HSN" },
    "source_separation": { "sampling_rate": 22050, "max_duration_s": 30 },
    ...
  }
}

To recreate a specific version follow the setup instructions on the GitHub repository.

Checkout the exact commit from metadata.json

git checkout <git_commit>

And run the pipeline as described on the GitHub repository.

Curation Rationale

Synthetic mixtures are used to enable precise control over polyphony level and SNR, which is not possible with naturally recorded soundscapes. This allows systematic evaluation of model performance across a range of acoustic conditions.

---

Known Limitations

• Mixtures are synthetic — real-world soundscapes may exhibit different acoustic properties and overlap patterns not captured here.
• Source recordings are subject to the biases and coverage gaps present in BirdSet and Xeno-canto (e.g. geographic and species imbalances).
• Currently limited to the HSN subset; further subsets will be added progressively.

Credits

This project builds on the following open-source work:

• BirdSet — Bird call audio data
• Perch — BirdMixIT source separation model
• BirdnetLib - Python api for BirdNET-Analyzer
• BirdNET Analyzer - Bird call analyzer
• TU Studio HPC Cluster ML Workflow — SLURM pipeline architecture

Citations

BirdSet: Rauch et al., 2024. BirdSet: A Large-Scale Dataset for Audio Classification in Avian Bioacoustics. arXiv:2403.10380.

BibTeX

bibtex@misc{rauch2024birdset, title={BirdSet: A Large-Scale Dataset for Audio Classification in Avian Bioacoustics}, author={Lukas Rauch et al.}, year={2024}, eprint={2403.10380}, archivePrefix={arXiv}, url={https://arxiv.org/abs/2403.10380}, }

eBird Taxonomy: Clements et al., 2024. The eBird/Clements Checklist of Birds of the World: v2024. Cornell Lab of Ornithology.

BibTeX

bibtex@misc{clements2024ebird, title={The eBird/Clements Checklist of Birds of the World: v2024}, author={Clements, J. F. and Rasmussen, P. C. and Schulenberg, T. S. and Iliff, M. J. and Fredericks, T. A. and Gerbracht, J. A. and Lepage, D. and Spencer, A. and Billerman, S. M. and Sullivan, B. L. and Smith, M. and Wood, C. L.}, year={2024}, howpublished={\url{https://www.birds.cornell.edu/clementschecklist/download/}} }

Bird-MixIT: Denton et al., 2021. Improving Bird Classification with Unsupervised Sound Separation. arXiv:2110.03209.

BibTeX

bibtex@misc{denton2021improvingbirdclassificationunsupervised, title={Improving Bird Classification with Unsupervised Sound Separation}, author={Tom Denton and Scott Wisdom and John R. Hershey}, year={2021}, eprint={2110.03209}, archivePrefix={arXiv}, primaryClass={eess.AS}, url={https://arxiv.org/abs/2110.03209}, }

MixIT: Wisdom et al., 2020. Unsupervised Sound Separation Using Mixture Invariant Training. arXiv:2006.12701.

BibTeX

bibtex@misc{wisdom2020unsupervisedsoundseparationusing, title={Unsupervised Sound Separation Using Mixture Invariant Training}, author={Scott Wisdom and Efthymios Tzinis and Hakan Erdogan and Ron J. Weiss and Kevin Wilson and John R. Hershey}, year={2020}, eprint={2006.12701}, archivePrefix={arXiv}, primaryClass={eess.AS}, url={https://arxiv.org/abs/2006.12701}, }

BirdNET: Kahl et al., 2021. BirdNET: A deep learning solution for avian diversity monitoring. Ecological Informatics, 61.

BibTeX

bibtex@article{kahl2021birdnet, title={BirdNET: A deep learning solution for avian diversity monitoring}, author={Kahl, Stefan and Wood, Connor M and Eibl, Maximilian and Klinck, Holger}, journal={Ecological Informatics}, volume={61}, pages={101236}, year={2021}, publisher={Elsevier} }

---

License

This dataset is released under CC BY-NC 4.0. Source recordings inherit the licensing terms of BirdSet — please consult the BirdSet dataset card for details.

---