Card: add western-palearctic tile row

6642faf verified 29 days ago

8.42 kB

pretty_name: Perch v2 Models (full + regional catalog)
license: apache-2.0
base_model:
  - cgeorgiaw/Perch
tags:
  - bioacoustics
  - birdnet
  - perch
  - bird
  - vocalization
  - audio-classification
  - onnx
library_name: onnx

Perch v2 Models (full + regional catalog)

Google's Perch v2 bioacoustic classifier in three deployment variants, plus a catalog of region-specific slices that are smaller and faster while staying numerically identical (bit-exact) to the full model on the species they keep.

Origin and attribution

Perch v2 by Google Research (bird-vocalization-classifier): EfficientNet-B3, ~12M embedding + ~91M classification params, ~15,000 species.
ONNX conversion and the DFT-to-MatMul (no_dft) optimization by justinchuby.
Labels from cgeorgiaw/Perch (iNaturalist taxonomy).
Regional slicing uses the BirdNET Geomodel v3.0 range filter (birdnet-team/geomodel) to pick each region's species.

Variants and hardware

filename token	precision	best for
`_fp32`	FP32, with DFT	GPU (CUDA/TensorRT), Intel CPU
`_no_dft_fp32`	FP32, DFT removed	OpenVINO (RPi5 fast path); also runs on ORT/CUDA
`_int8_arm`	partial INT8 (MatMul-only)	ARM CPU / Raspberry Pi, low RAM

Full model

file	MB
`full/perch_v2_fp32.onnx`	409
`full/perch_v2_no_dft_fp32.onnx`	413
`full/perch_v2_int8_arm.onnx`	131
`full/perch_v2_labels.txt`	14,795 classes

Why regional models

These slices are built for real-time detection on resource-constrained devices, phones, Raspberry Pi and other single-board computers, where running the full 14,795-class Perch v2 continuously is costly in both RAM and CPU. Most of that cost goes to recognising species that cannot occur at the listener's location. Restricting the model to a region's species shrinks the memory footprint and the per-inference compute (the classifier head is ~88% of the model), so an always-on detector keeps up with the live audio stream and leaves headroom for the rest of the application, on hardware where the full model would struggle. Each tile stays bit-exact to the full model on the species it keeps; the only change is that out-of-region species are not emitted, which at a fixed monitoring location is exactly what you want.

Regional catalog

Each tile: BirdNET Geomodel v3.0 range filter (top ~800 species for temperate regions, ~1200 for bird-rich tropical/subtropical ones, up to ~3500 for the hyper-diverse Neotropics) + 198 FSD50K sound events + a 27-species cosmopolitan core. Ships _no_dft_fp32 (OpenVINO/GPU) and _int8_arm (ARM) + labels + indices. All bit-exact vs the full model on the species they keep.

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. regional/groups.json lists the same grouping (ordered continents -> tiles) in one file, and each tile's metadata.json carries its group / group_display / group_order, so an application can rebuild these sections without scraping this table.

Europe

region	classes	fp32 MB	int8-arm MB
`nordic`	638	65.0	44.5
`british-isles`	776	68.4	45.4
`central-europe`	873	70.8	46.0
`baltics`	655	65.4	44.6
`iberia`	856	70.4	45.9
`southern-europe`	839	70.0	45.8
`eastern-europe`	739	67.5	45.2
`western-palearctic`	1599	88.7	50.5
`iceland`	591	63.9	44.3
`svalbard`	480	61.1	43.6
`canary-islands`	598	64.0	44.3
`madeira`	459	60.6	43.4
`azores`	426	59.8	43.2

Asia

region	classes	fp32 MB	int8-arm MB
`south-asia-peninsular`	879	70.9	46.0
`indo-gangetic`	1406	83.9	49.3
`himalaya`	1407	83.9	49.3
`japan`	799	69.0	45.5
`china-northeast`	877	70.9	46.0
`china-north-central`	976	73.3	46.6
`china-southeast`	1373	83.1	49.1
`china-southwest`	1409	84.0	49.3
`tibet`	1407	83.9	49.3

North America

region	coverage	classes	fp32 MB	int8-arm MB
`north-america-east`		999	73.9	46.8
`north-america-west`		1002	74.0	46.8

South America

region	classes	fp32 MB	int8-arm MB
`amazonia`	3388	132.7	61.5
`andes`	3535	136.3	62.4
`eastern-brazil`	2184	103.0	54.1
`southern-cone`	1855	95.0	52.0
`galapagos`	336	57.6	42.7

Africa

region	classes	fp32 MB	int8-arm MB
`southern-africa`	1002	74.0	46.8
`reunion`	274	56.1	42.3
`mauritius`	272	56.0	42.3
`seychelles`	318	57.2	42.6
`cape-verde`	346	57.8	42.7
`sao-tome-principe`	324	57.3	42.6

Oceania

region	classes	fp32 MB	int8-arm MB
`australia-east`	946	72.6	46.4
`new-zealand`	486	61.3	43.6
`hawaii`	478	61.1	43.6
`new-caledonia`	377	58.6	42.9

Usage

Each model takes 5 s of 32 kHz mono audio ([1, 160000]) and outputs a label vector of logits over its species list; pair it with the sibling *_labels.txt (line count matches the logit count). Pick a variant by hardware (table above). Confidence is a softmax over the model's own classes, so a regional tile normalizes over fewer species than the full model; recalibrate detection thresholds per model.

Provenance

Regional slices are gathered from the ProtoPNet head of perch_v2_no_dft.onnx and validated bit-exact against the full model on the species they keep. Perch v2 is by Google; see the license above.