BirdNET+ V3.0 (developer preview 3.1) 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.

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: EfficientNetV2-S backbone, three-head classifier, 11,560 species. Base preview models are distributed by the BirdNET team on Zenodo.

The GPU-native conversion and the regional slicing in this repo are derived work under CC BY-SA 4.0.

The BirdNET team's ONNX export computes its log-mel front-end with an in-graph ONNX STFT op. ONNX Runtime has no CUDA/TensorRT kernel for STFT, and OpenVINO cannot translate it at all, so the stock export cannot run on GPU or OpenVINO. In these models the STFT is replaced with a mathematically equivalent Conv1d (framing + windowed DFT), so the graph is fully supported on ONNX Runtime CPU, OpenVINO (x86 + Intel GPU), CUDA, and TensorRT. On multi-threaded CPU it is also slightly faster than the stock model (ORT's STFT op is poorly parallelized). 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, 11,560] and embeddings [batch, 1280].

FP16 is a download-size lever, not a CPU one. ONNX Runtime upconverts FP16 weights at inference, so on CPU (x86 and ARM alike) FP16 is equal to or slower than FP32, and on a Raspberry Pi 4 it uses more RAM than FP32. For low-RAM CPU devices the right choice is a regional FP32 model, not FP16.

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.

OpenVINO is ~2.2x faster than ONNX Runtime on a Raspberry Pi 5. Batching helps only on a GPU (~1.5x at batch 8, FP16); on CPU use batch 1.

The three-head classifier is ~80% of the model's parameters, so restricting it to a region's species set cuts memory a lot (~67% less RAM: 685 MB -> 229 MB on a Raspberry Pi 5) and latency modestly (15-30%). The parameters live in the heads, but the convolutional backbone still does most of the arithmetic, so the headline win is footprint rather than speed. That is exactly what an always-on detector on a phone or single-board computer needs. 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 range filter (top ~800 by occurrence probability + a cosmopolitan core), using the same tiles as the Perch v2 regional catalog.

Each tile ships _fp32.onnx (OpenVINO/GPU/CPU) and _fp16.onnx + labels + indices. All are numerically identical to the full model on the species they keep (max abs difference < 1e-6); species sets come from the BirdNET Geomodel v3.0 range filter, using the same tiles as the Perch v2 regional catalog.

Each tile folder also has coverage.png (a map of the region it covers) and metadata.json (species count, covered countries, and the continental group it belongs to).

Tiles are organised by continent below, in the same order as the Perch v2 catalog. regional/groups.json lists the continent order, and each tile's metadata.json carries its group / group_display / group_order, so an application can rebuild these sections without scraping this table.

Files are named birdnet-<upstream>-[<region>-]<variant>-b<build>.onnx, where upstream is the BirdNET release these weights come from (v3.0-preview3.1) and build is this repo's build revision of that conversion. A new upstream release or a new conversion always gets a new filename; a published file's bytes never change. models.json is the machine-readable index: it lists every file with its variant, checksum, input contract, and which backends it actually runs on (including where a backend is not supported), plus a selection map from hardware/backend to the file to download.