BirdNET+ V3.0 (GPU-native full + regional catalog)

BirdNET+ V3.0.1 acoustic classifier, re-exported to a GPU-native ONNX graph, plus a catalog of region-specific slices that are much smaller while staying numerically identical to the full model on the species they keep.

Licence and terms

Each release came with its own upstream terms, listed here with its licence. The repo-level licence tag is the current release's.

  • BirdNET+ V3.0.1 (birdnet-plus-v3.0.1), Apache-2.0: the Apache-2.0 model licence and acceptable-use guidance of the BirdNET+ V3.0.1 release (Zenodo, https://doi.org/10.5281/zenodo.24362980), LICENSE-v3.0.1.md and ACCEPTABLE_USE-v3.0.1.md at the repo root, unchanged from upstream.
  • BirdNET+ V3.0 (birdnet-plus-v3.0), unlisted: the Apache-2.0 model licence and acceptable-use guidance of the BirdNET+ V3.0 release (Zenodo, https://doi.org/10.5281/zenodo.24162518), LICENSE-v3.0.md and ACCEPTABLE_USE-v3.0.md at the repo root, unchanged from upstream.

Origin and attribution

Powered by BirdNET (https://birdnet.cornell.edu/). BirdNET+ V3.0 is developed by the K. Lisa Yang Center for Conservation Bioacoustics (Cornell University), Chemnitz University of Technology, and the Museum fuer Naturkunde Berlin. It builds on Kahl et al. (2021) and Lasseck (2018). Architecture of this release: EfficientNetV2-S backbone, three-head classifier, 11,725 classes.

The GPU-native conversion and the regional slicing in this repo are derived work under Apache-2.0.

Releases in this repo

The current release is BirdNET+ V3.0.1 (birdnet-plus-v3.0.1).

Unlisted releases:

  • BirdNET+ V3.0 (birdnet-plus-v3.0), unlisted on 2026-11-20: superseded by 3.0.1 in the same repo; its front end returns a wrong result on a constant input at 0.9 of full scale or more under ONNX Runtime, and it is slower; kept for A/B runs only. Its files, its SHA256SUMS lines and its licence texts stay unchanged; no models.json lists it.

Changelog

Build 3 (2026-11-21)

Adds a regional model for Central India, where the existing Indian regional models do not fit.

  • New regional model for Central India: the Deccan plateau and the Narmada and Godavari basins, roughly Madhya Pradesh, Maharashtra, Chhattisgarh and Telangana.
  • It covers 656 wild species of the region, for example Indian Roller, Sarus Crane, Indian Peafowl and Painted Spurfowl, and it includes the domestic dog and all other domestic animals and sound classes like every other regional model.
  • All other regional models and the global model are unchanged.

Build 2 (2026-11-21)

Every regional model now includes the domestic dog and all other domestic animals, which some regions were missing.

  • Every regional model now includes the domestic dog, cat, horse, cattle, sheep, goat, chicken and pigeon, and other pets and farmed animals such as rabbit, donkey, goldfish, carp, budgerigar and canary.
  • Before this build the domestic dog was missing from 17 regional models, including the Nordic one, so barking dogs could not be reported there. Cat, horse, cattle, sheep, goat and chicken were missing from many regions as well.
  • Six regions with a species limit gained between 4 and 24 more wild species each, for example Corn Crake and Asian Elephant in the Himalaya region and Fieldfare in western North America.
  • Regional models grow by 5 to 28 classes and no regional model lost a class.
  • The global model and the sound classes are unchanged.

What "GPU-native" means

The log-mel front-end of these models is plain tensor arithmetic: the audio is framed by reshaping, each frame has its mean removed before the Hann window, a two-stage DFT runs as MatMuls on constant matrices, and the mel matrix is applied as a MatMul or Einsum. The graph holds no STFT or DFT op, which ONNX Runtime lacks on CUDA and TensorRT and OpenVINO cannot translate. Every file was run on ONNX Runtime CPU (x86-64 and a Raspberry Pi 5) and on OpenVINO at f32 (CPU on both, and two Intel integrated GPUs), and its scores were compared with the x86-64 ONNX Runtime scores of the same file; the manifest's supported and verified_on follow those runs. OpenVINO must run at f32: its default f16 hint gives wrong scores on the Cortex-A76 and on Intel GPUs. On the Intel GPUs it is the faster choice (63 ms against 97 ms for ONNX Runtime CPU for the global model on an Intel Iris Plus Graphics 640) and the manifest recommends it there; on CPUs ONNX Runtime is recommended (on a Raspberry Pi 5 OpenVINO CPU is slower on the global model, 290 against 260 ms, and about the same on the nordic tile, 202 against 218 ms, at about 1.8 times the memory). A transient error of up to 5e-2 on isolated inputs was seen in a few runs on an Intel Iris Plus Graphics 640. The CUDA and TensorRT entries were not run for this release. One model serves CPU and GPU; there is no separate CPU/GPU build.

Input: raw audio [batch, 160000] (5 s @ 32 kHz, matching Perch v2 so a host can feed both classifiers the same buffer). Outputs: predictions [batch, 11,725] (a static dimension, sigmoid already applied) and embeddings [batch, 1280].

Variants and hardware

precision best for
-fp32- run this: ONNX Runtime and OpenVINO (BirdNET-Go runs v3.0 at f32 on every OpenVINO device)

This release ships FP32 only. For low-RAM CPU devices the right choice is a regional FP32 model.

int8-arm is not shipped for v3.0: the backbone is Conv-dominated and int8-arm only quantizes MatMul/Gemm weights, so it barely helps here.

Full model

file classes MB
full/birdnet-plus-v3.0.1-fp32-b1.onnx 11,725 345
full/birdnet-plus-v3.0.1-labels-b1.txt 11,725 -

Why regional models

The three-head classifier holds a large share of the model's parameters, so restricting it to a region's species set cuts memory and download size. Each tile is numerically identical to the full model on the species it keeps (max abs difference vs the full model < 1e-6); the only change is that out-of-region species are not emitted, which at a fixed monitoring location is exactly what you want. Region species sets come from the BirdNET Geomodel v3.0.2 range filter (species by occurrence probability, up to a per-tile cap, plus a cosmopolitan core), using the same tiles as the Perch v2 regional catalog. The cap is 800 for most temperate tiles and higher for species-rich ones. Every tile also contains Homo sapiens and the release's non-animal sound classes, whatever its range filter says.

A species the geomodel names differently from the classifier (a moved genus, a changed spelling) is matched through a reviewed rename table, so the tiles keep species an exact name match would have dropped.

Regional catalog

Each tile ships an fp32 model + labels + indices + coverage.png + metadata. regional/groups.json lists the continent grouping.

Europe

region coverage classes build fp32 MB
azores 263 2 97.9
baltics 479 2 102.5
british-isles 590 2 104.9
canary-islands 428 2 101.5
central-europe 673 2 106.7
eastern-europe 556 2 104.2
iberia 654 2 106.3
iceland 429 2 101.5
madeira 293 2 98.5
nordic 464 2 102.2
southern-europe 653 2 106.3
svalbard 323 2 99.2
western-palearctic 1446 2 123.4

Asia

region coverage classes build fp32 MB
central-india 708 3 107.5
china-north-central 809 2 109.7
china-northeast 706 2 107.4
china-southeast 1191 2 117.9
china-southwest 1256 2 119.3
himalaya 1254 2 119.2
indo-gangetic 1212 2 118.3
japan 629 2 105.8
south-asia-peninsular 690 2 107.1
tibet 1239 2 118.9

North America

region coverage classes build fp32 MB
canada-alaska 793 2 109.3
north-america-east 846 2 110.4
north-america-west 847 2 110.5

South America

region coverage classes build fp32 MB
amazonia 3196 2 161.0
andes 3325 2 163.8
eastern-brazil 1986 2 135.0
galapagos 181 2 96.1
southern-cone 1648 2 127.7

Africa

region coverage classes build fp32 MB
cape-verde 187 2 96.3
central-africa 1556 2 125.7
east-africa 1840 2 131.8
mauritius 116 2 94.7
reunion 118 2 94.8
sao-tome-principe 159 2 95.7
seychelles 162 2 95.7
southern-africa 847 2 110.5
west-africa 1043 2 114.7

Oceania

region coverage classes build fp32 MB
australia-east 766 2 108.7
hawaii 304 2 98.8
new-caledonia 214 2 96.8
new-zealand 302 2 98.7

Deployment

  • Point your host at a tile's model + its labels file (matching counts) and restart. Labels are Scientific name_Common name.
  • Confidence is a per-class sigmoid; slicing does not change a kept class's score (unlike a softmax model), so thresholds transfer directly from the full model.
  • Binaries are hosted here on HuggingFace; the derivation tooling lives separately.
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support