BirdNET+ V3.0 model files
Authors/Creators
Description
BirdNET+ V3.0 Global 12K (Nov 2026) identifies animal sounds in audio recordings. It recognizes 11,709 animal species, most of them birds, and 16 non-animal sound classes.
Version 3.0.1 (Nov 2026):- The V3.0 models were exported again with conversion fixes. Weights, classes and labels are unchanged.
- Silent (all-zero) input no longer produces NaN in the ONNX and FP16 TFLite models.
- The FP16 ONNX models are now mixed precision: same size, within about 0.01 of the FP32 scores, and safe from float16 overflow on GPUs and mobile devices. Their second output is now named embeddings (was embeddings_out), as in the FP32 model.
- The spectrogram is computed without an STFT operator, for wider runtime support.
- The TensorFlow SavedModel accepts only its export shape (1 x 96,000 samples, 3 s) and raises an error for other shapes.
- New in this record: the BirdNET+ Geomodel V3.0.5, the BirdNET+ Taxonomy v0.5, and the two models bundled with the BirdNET Live app.
Species and classes:
- 9,788 birds, 802 insects, 649 amphibians, 419 mammals (including humans), 35 fishes, 15 reptiles, and 1 arachnid
- 16 non-animal classes like gun, chainsaw, siren, etc.
- Curated class list: one class per species, and classes with too little training data removed
- Labels linked to the BirdNET+ Taxonomy v0.5 (Nov 2026): BirdNET ID, scientific name, iNaturalist ID, eBird code, and common names in 36 languages
Training data from large public archives, collected up to:
- Macaulay Library: 17 Mar 2026
- iNaturalist: 12 Jan 2026
- Xeno-canto: 6 Jan 2026
- Animal Sound Archive Berlin (Tierstimmenarchiv): 23 Apr 2025
- Observation.org: 28 Dec 2024
Training data also included non-public collections (e.g., validated snippets from collaborators).
Key changes vs. BirdNET V2.4:
- Variable-length input instead of fixed 3-second segments
- 32 kHz audio input instead of 48 kHz
- Improved architecture and training procedure
- Much larger and more diverse training dataset
- Nearly twice as many classes (V2.4: 6,522), with many more non-bird species
- About twice the average precision on WABAD soundscapes
Key changes vs. Developer Preview 3.1 (11K, Jun 2026):
- Larger, curated species list
- First non-animal classes for human-made and environmental sounds
- Labels linked to the BirdNET+ Taxonomy, with translated common names
- New Core ML export for iOS and macOS
Evaluation on WABAD:
WABAD (Pérez-Granados et al. 2025, https://zenodo.org/records/14191524) is a public set of expert-annotated soundscapes. We scored 66 sites on six continents, with 88,967 annotated calls of about 1,000 bird species. Average precision (AP) over all 3-second windows, on the species all models can predict (Nov 2026):
- BirdNET+ V3.0: AP 0.552
- BirdNET V2.4: AP 0.272
Input and output:
- Mono audio at 32 kHz, variable length. TFLite models are most accurate on 3-second segments; Core ML models and the TensorFlow SavedModel take exactly 3 s.
- Outputs: a confidence score per class and an embedding vector
Audio model files (all names start with BirdNET+_V3.0_Global_12K):
- _FP32.pt: TorchScript (PyTorch)
- _FP32.onnx, _FP16.onnx: ONNX
- _FP32.tflite, _FP16.tflite: LiteRT (TensorFlow Lite)
- _FP32_Protobuf.zip: TensorFlow SavedModel
- _FP32.mlpackage.zip, _FP16.mlpackage.zip: Core ML (iOS 16 / macOS 13 or later)
- _FP16_pruned.onnx, _FP16_pruned.tflite: smaller models (~68 MB) for mobile and edge devices; they use one of the model's three classifier heads
- _Labels.csv, _Labels.xlsx: one row per model output, in the same order for all files above
- _Source_Checkpoint.zip: the unpruned source model (BiophonyNET_26.10.1) with all 16,852 training classes (TorchScript and ONNX, FP32) and its labels, for research
FP16 models are half the size of FP32 with minimal accuracy loss.
BirdNET+ Geomodel (BirdNET+_Geomodel_V3.0.5_Global_14K.zip):
The geomodel predicts which species are likely to occur at a given location in a given week of the year. Use it to filter the audio model's detections by place and season.
- Input: latitude, longitude and week (1 to 48). Output: one occurrence probability per class.
- 14,057 classes. V3.0.5 has the trained weights of the Geomodel V3.0.4 release, with the classes reduced to those in the BirdNET+ Taxonomy and labelled with BirdNET IDs, so its output joins to the audio model's labels by BirdNET ID.
- The zip contains TorchScript, ONNX (FP32, FP16), LiteRT (FP32, FP16), TensorFlow SavedModel, Core ML (FP32, FP16), the PyTorch checkpoint, and the labels (_Labels.csv, _Labels.txt). The TorchScript model returns logits, as in earlier geomodel releases; apply a sigmoid.
- Training code and documentation: https://github.com/birdnet-team/geomodel
BirdNET+ Taxonomy (BirdNET+_Taxonomy_v0.5-Nov2026.zip):
The species list behind the labels of both models, as CSV and JSON: 18,093 records (species and non-species sound classes) with BirdNET IDs, scientific names, common names in 36 languages, and external IDs such as iNaturalist, eBird and GBIF. The JSON also has descriptions and image links. Browse it at https://birdnet.cornell.edu/taxonomy/
BirdNET Live models (BirdNET+_V3.0_Global_10K_BNL.zip):
The two models the BirdNET Live app (https://github.com/birdnet-team/birdnet-live-app) runs on-device, packaged for use in other applications. Both are ONNX FP16 and share one species set: the audio model keeps a species only if the geomodel can predict it somewhere in the world.
- BirdNET+_V3.0_Global_10K_FP16_BNL.onnx: audio model (68 MB), 10,064 classes: 10,048 species (birds, amphibians, insects, mammals) and the 16 non-animal sound classes. It uses one of the model's three classifier heads, like the _FP16_pruned models above, and gives the same scores as _FP16_pruned.onnx for the classes it keeps.
- BirdNET+_Geomodel_V3.0.5_Global_10K_FP16_BNL.onnx: geomodel (7 MB) for the same 10,048 species, with the same probabilities as the FP16 ONNX geomodel above.
- BirdNET+_V3.0_Global_10K_Labels.csv: one labels file for both models. One row per audio class with BirdNET ID, scientific name, common names in 36 languages, and geo_ix, the index of the species in the geomodel output (empty for sound classes).
- README.md with input and output details and a code example, the model license, the acceptable use policy, and SHA256SUMS.
Files
ACCEPTABLE_USE.txt
Files
(3.4 GB)
| Name | Size | |
|---|---|---|
|
md5:92f52349d7a08f74167f096702bac26a
|
3.0 kB | Preview Download |
|
md5:d30b7f51ad301889ab193a384c72f1a1
|
111.2 MB | Preview Download |
|
md5:0310312e852326e17e45e5b2dcf64655
|
69.8 MB | Preview Download |
|
md5:b2ee809cbb67c0988afcbd2d156d4fc9
|
69.6 MB | Preview Download |
|
md5:e2cc0ee961da3f52aee2428c5326f110
|
156.5 MB | Preview Download |
|
md5:3f284bdeae7ffb38718c391839b58b60
|
173.9 MB | Download |
|
md5:f6e0ac4777fe9cb9f3fc20bb325e89fe
|
172.3 MB | Download |
|
md5:6b42d8905f25f856fb8ad91544138914
|
72.3 MB | Download |
|
md5:a3575f46173acbe089cabe076b85998b
|
70.9 MB | Download |
|
md5:0e5d3af438cf7fbfe3c59f65c37e10a0
|
316.6 MB | Preview Download |
|
md5:07c6b4948d7b7bd1e963dcbd03aa7afa
|
345.4 MB | Download |
|
md5:70aebf9efba6e1c8d58036c88007471b
|
345.2 MB | Download |
|
md5:be95f36ca540f722663697b3346a1ed8
|
344.1 MB | Download |
|
md5:16c99f4fd9a617922c62732e5c2e28e4
|
317.4 MB | Preview Download |
|
md5:bb25847af8a581e1382081382f01e762
|
8.0 MB | Preview Download |
|
md5:725ce93002b58ac11e1dce42b457f783
|
4.7 MB | Download |
|
md5:ee712a55d11e32a8fc4ed2e294a4ac48
|
837.3 MB | Preview Download |
|
md5:77b299ba50d3119cd28c5a2b6fc0ff4b
|
11.8 kB | Preview Download |
Additional details
Software
- Repository URL
- https://github.com/birdnet-team/birdnet-V3.0-dev
- Development Status
- Active