checkpoints/README.md

BirdNET Analyzer Model Version History

This repository contains only the latest version of the BirdNET-Analyzer model. We will track the changes in this document and also provide permanent download links to previous versions for testing and/or other use cases.

You can download and unzip previous model versions, make sure to update the MODEL_PATH and LABELS_FILE variable in the config.py file to the version you want to use.

Model update history:

V2.4

• more than 6,000 species worldwide
• covers frequencies from 0 Hz to 15 kHz with two-channel spectrogram (one for low and one for high frequencies)
• 0.826 GFLOPs, 50.5 MB as FP32
• enhanced and optimized metadata model
• global selection of species (birds and non-birds) with 6,522 classes (incl. 10 non-event classes)
• Download here: BirdNET-Analyzer-V2.4.zip

V2.3

• slightly larger (36.4 MB vs. 21.3 MB as FP32) but smaller computational footprint (0.698 vs. 1.31 GFLOPs) than V2.2
• larger embedding size (1024 vs 320) than V2.2 (hence the bigger model)
• enhanced and optimized metadata model
• global selection of species (birds and non-birds) with 3,337 classes (incl. 10 non-event classes)
• Download here: BirdNET-Analyzer-V2.3.zip

V2.2

• smaller (21.3 MB vs. 29.5 MB as FP32) and faster (1.31 vs 2.03 GFLOPs) than V2.1
• maintains same accuracy as V2.1 despite more classes
• global selection of species (birds and non-birds) with 3,337 classes (incl. 10 non-event classes)
• Download here: BirdNET-Analyzer-V2.2.zip

V2.1

• same model architecture as V2.0
• extended 2022 training data
• global selection of species (birds and non-birds) with 2,434 classes (incl. 10 non-event classes)
• Download here: BirdNET-Analyzer-V2.1.zip

V2.0

• same model design as 1.4 but a bit wider
• extended 2022 training data
• global selection of species (birds and non-birds) with 1,328 classes (incl. 10 non-event classes)
• Download here: BirdNET-Analyzer-V2.0.zip

V1.4

• smaller, deeper, faster
• only 30% of the size of V1.3
• still linear spectrogram and EfficientNet blocks
• extended 2021 training data
• 1,133 birds and non-birds for North America and Europe
• Download here: BirdNET-Analyzer-V1.4.zip

V1.3

• Model uses linear frequency scale for spectrograms
• uses V2 fusion blocks and V1 efficient blocks
• extended 2021 training data
• 1,133 birds and non-birds for North America and Europe
• Download here: BirdNET-Analyzer-V1.3.zip

V1.2

• Model based on EfficientNet V2 blocks
• uses V2 fusion blocks and V1 efficient blocks
• extended 2021 training data
• 1,133 birds and non-birds for North America and Europe
• Download here: BirdNET-Analyzer-V1.2.zip

V1.1

• Model based on Wide-ResNet (aka "App model")
• extended 2021 training data
• 1,133 birds and non-birds for North America and Europe
• Download here: BirdNET-Analyzer-V1.1.zip

App Model

• Model based on Wide-ResNet
• ~3,000 species worldwide
• currently deployed as BirdNET app model
• Download here: BirdNET-Analyzer-App-Model.zip