Skip to content

kahst/BirdNET-Analyzer

Repository files navigation

BirdNET-Analyzer

Automated scientific audio data processing and bird ID.

CC BY-NC-SA 4.0 Supported OS Number of species Downloads

Email Twitter Follow Subreddit subscribers

Introduction

This repo contains BirdNET models and scripts for processing large amounts of audio data or single audio files. This is the most advanced version of BirdNET for acoustic analyses and we will keep this repository up-to-date with new models and improved interfaces to enable scientists with no CS background to run the analysis.

Feel free to use BirdNET for your acoustic analyses and research. If you do, please cite as:

@article{kahl2021birdnet,
  title={BirdNET: A deep learning solution for avian diversity monitoring},
  author={Kahl, Stefan and Wood, Connor M and Eibl, Maximilian and Klinck, Holger},
  journal={Ecological Informatics},
  volume={61},
  pages={101236},
  year={2021},
  publisher={Elsevier}
}

About

Go to https://birdnet.cornell.edu to learn more about the project.

Want to use BirdNET to analyze a large dataset? Don’t hesitate to contact us: ccb-birdnet@cornell.edu

Follow us on Twitter @BirdNET_App

We also have a discussion forum on Reddit if you have a general question or just want to chat.

Have a question, remark, or feature request? Please start a new issue thread to let us know. Feel free to submit a pull request.

Contents

1. Model version update

V2.4, June 2023

  • 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)

You can find a list of previous versions here: BirdNET-Analyzer Model Version History

Species range model V2.4 - V2, Jan 2024

  • updated species range model based on eBird data

  • more accurate (spatial) species range prediction

  • slightly increased long-tail distribution in the temporal resolution

  • see this discussion post for more details

2. Technical Details

Model V2.4 uses the following settings:

  • 48 kHz sampling rate (we up- and downsample automatically and can deal with artifacts from lower sampling rates)

  • we compute 2 mel spectrograms as input for the convolutional neural network:

    • first one has fmin = 0 Hz and fmax = 3000; nfft = 2048; hop size = 278; 96 mel bins

    • second one has fmin = 500 Hz and fmax = 15 kHz; nfft = 1024; hop size = 280; 96 mel bins

  • both spectrograms have a final resolution of 96x511 pixels

  • raw audio will be normalized between -1 and 1 before spectrogram conversion

  • we use non-linear magnitude scaling as mentioned in Schlüter 2018

  • V2.4 uses an EfficienNetB0-like backbone with a final embedding size of 1024

  • See this comment for more details

3. Usage guide

This document provides instructions for downloading and installing the GUI, and conducting some of the most common types of analyses. Within the document, a link is provided to download example sound files that can be used for practice.

Download the PDF here: BirdNET-Analyzer Usage Guide

Watch our presentation on how to use BirdNET-Analyzer to train your own models: BirdNET - BioacousTalks at YouTube

4. Showroom

BirdNET powers a number of fantastic community projects dedicated to bird song identification, all of which use models from this repository. These are some highlights, make sure to check them out!

Table 1. Community projects
Project Description

HaikuBox

HaikuBox
Once connected to your WiFi, Haikubox will listen for birds 24/7. When BirdNET finds a match between its thousands of labeled sounds and the birdsong in your yard, it identifies the bird species and shares a three-second audio clip to the Haikubox website and smartphone app.

Learn more at: HaikuBox.com

BirdNET-PI

BirdNET-Pi
Built on the TFLite version of BirdNET, this project uses pre-built TFLite binaries for Raspberry Pi to run on-device sound analyses. It is able to recognize bird sounds from a USB sound card in realtime and share its data with the rest of the world.

Learn more at: BirdNETPi.com

BirdWeather

BirdWeather
This site was built to be a living library of bird vocalizations. Using the BirdNET artificial neural network, BirdWeather is continuously listening to over 1,000 active stations around the world in real-time.

Learn more at: BirdWeather.com

ecoPI:Bird

ecoPi:Bird
The ecoPi:Bird is a device for automated acoustic recordings of bird songs and calls, with a self-sufficient power supply. It facilitates economical long-term monitoring, implemented with minimal personal requirements.

Learn more at: oekofor.netlify.app

Chirpity

Chirpity
Discover the wonders of bird identification with Chirpity, a desktop application powered by cutting-edge Machine Learning. With the option to choose between BirdNET or the native Chirpity model, finely tuned for Nocturnal Flight Calls, you have the flexibility to tailor your analysis to your specific needs. Perfect for enthusiasts and researchers alike, Chirpity is particularly well-suited for Nocmig and other extensive field recordings. Chirpity is available on both Windows and Mac platforms.

Learn more at: chirpity.mattkirkland.co.uk

Go-BirdNET

Go-BirdNET
Go-BirdNET is an application inspired by BirdNET-Analyzer. While the original BirdNET is based on Python, Go-BirdNET is built using Golang, aiming for simplified deployment across multiple platforms, from Windows PCs to single board computers like Raspberry Pi.

Learn more at: github.com/tphakala/go-birdnet

whoBIRD

whoBIRD
whoBIRD empowers you to identify birds anywhere, anytime, without an internet connection. Built upon the TFLite version of BirdNET, this Android application harnesses the power of machine learning to recognize birds directly on your device.

Learn more at: whoBIRD

Other cool projects:

Working on a cool project that uses BirdNET? Let us know and we can feature your project here.

5. Projects map

We have created an interactive map of projects that use BirdNET. If you are working on a project that uses BirdNET, please let us know here and we can add it to the map.

You can access the map here: Open projects map

6. Setup

6.1. Setup (Raven Pro)

If you want to analyze audio files without any additional coding or package install, you can now use Raven Pro software to run BirdNET models. After download, BirdNET is available through the new "Learning detector" feature in Raven Pro. For more information on how to use this feature, please visit the Raven Pro Knowledge Base.

Download the newest model version here, extract the zip-file and move the extracted folder to the Raven models folder. On Windows, the models folder is C:\Users\<Your user name>\Raven Pro 1.6\Models. Start Raven Pro and select BirdNET_GLOBAL_6K_V2.4_Model_Raven as learning detector.

6.2. Setup (birdnetlib)

The easiest way to setup BirdNET on your machine is to install birdnetlib through pip with:

pip3 install birdnetlib

Make sure to install Tensorflow Lite, librosa and ffmpeg like mentioned below. You can run BirdNET with:

from birdnetlib import Recording
from birdnetlib.analyzer import Analyzer
from datetime import datetime

# Load and initialize the BirdNET-Analyzer models.
analyzer = Analyzer()

recording = Recording(
    analyzer,
    "sample.mp3",
    lat=35.4244,
    lon=-120.7463,
    date=datetime(year=2022, month=5, day=10), # use date or week_48
    min_conf=0.25,
)
recording.analyze()
print(recording.detections)

For more examples and documentation, make sure to visit pypi.org/project/birdnetlib/. For any feature request or questions regarding birdnetlib, please contact Joe Weiss or add an issue or PR at github.com/joeweiss/birdnetlib.

6.3. Setup (Ubuntu)

Install Python 3:

sudo apt-get update
sudo apt-get install python3-dev python3-pip
pip3 install --upgrade pip

Install TFLite runtime (recommended) or Tensorflow (has to be 2.15):

pip3 install tflite-runtime

OR

pip3 install tensorflow==2.15.0

Install Librosa to handle audio files:

pip3 install librosa resampy
sudo apt-get install ffmpeg

Clone the repository

git clone https://github.com/kahst/BirdNET-Analyzer.git
cd BirdNET-Analyzer

6.4. Setup (Windows)

Before you attempt to setup BirdNET-Analyzer on your Windows machine, please consider downloading our fully-packaged version that does not require you to install any additional packages and can be run "as-is".

You can download this version here: BirdNET-Analyzer Windows

  1. Download the BirdNET-Analyzer-setup.exe file

  2. Before installing, make sure to right-click the exe-file, select "Properties" and check the box "Unblock" under "Security" at the bottom of the "General" tab.

    • If Windows does not display this option, the file can be unblocked with the PowerShell 7 command Unblock-File -Path .\BirdNET-Analyzer.zip

  3. During installation, you may see a warning "Windows protected your PC" due to the lack of a digital signature. Simply select "More info" and then "Run anyway" to proceed with the installation.

  4. Follow the on-screen instructions

  5. After installation, click the desktop icon or navigate to the extracted folder at C:\Users\<Your user name>\AppData\Local\Programs\BirdNET-Analyzer

  6. You can start the analysis through the command prompt with BirdNET-Analyzer.exe --i "path\to\folder" ... (see Usage (CLI) section for more details), or you can launch BirdNET-Analyzer-GUI.exe to start the analysis through a basic GUI.

For more advanced use cases (e.g., hosting your own API server), follow these steps to set up BirdNET-Analyzer on your Windows machine:

Install Python 3.9 or higher (has to be 64bit version)

⚠️
Make sure to check: ☑ "Add path to environment variables" during install

Install Tensorflow (has to be 2.5 or later), Librosa and NumPy

  • Open command prompt with Win + S type "command" and click on "Command Prompt"

  • Type pip install --upgrade pip

  • Type pip install librosa resampy

  • Install Tensorflow by typing pip install tensorflow

ℹ️
You might need to run the command prompt as "administrator". Type Win + S, search for command prompt and then right-click, select "Run as administrator".

Install Visual Studio Code (optional)

Install BirdNET using Git (for simple download see below)

  • Download and install Git Bash: Download Git Bash installer

  • Select Visual Studio Code as default editor (optional)

  • Keep all other settings as recommended

  • Create folder in personal directory called "Code" (or similar)

  • Change to folder and right click, launch "Git bash here"

  • Type git clone https://github.com/kahst/BirdNET-Analyzer.git

  • Keep BirdNET updated by running git pull for BirdNET-Analyzer folder occasionally

Install BirdNET from zip

  • Download BirdNET: Download BirdNET Zip-file

  • Unpack zip file (e.g., in personal directory)

  • Keep BirdNET updated by re-downloading the zip file occasionally and overwrite existing files

Run BirdNET from command line

  • Open command prompt with Win + S type "command" and click on "Command Prompt"

  • Navigate to the folder where you installed BirdNET (cd path\to\BirdNET-Analyzer)

  • See Usage (CLI) section for command line arguments

ℹ️
With Visual Studio Code installed, you can right-click the BirdNET-Analyzer folder and select "Open with Code". With proper extensions installed (View → Extensions → Python) you will be able to run all scripts from within VS Code.

6.5. Setup (macOS)

ℹ️
Installation was only tested on M1 and M2 chips. Feedback on older Intel CPUs or newer M3 chips is welcome!

Requirements

You need to install the Xcode command-line tools:

xcode-select --install

Clone the git repository into your prefered folder if you have not done that yet:

cd ~/YOUR_FOLDER
git clone https://github.com/kahst/BirdNET-Analyzer.git
cd BirdNET-Analyzer

Setup the environment

We are going to create a virtual environment to install the required packages. Virtual environments allow you to manage separate package installations for different projects.

python3 -m venv venv-birdnet
source venv-birdnet/bin/activate
python -m pip install -U pip

The nexttime you want to use BirdNET, go to the BirdNET-Analyzer folder and run source venv-birdnet/bin/activate to activate the virtual environment.

Install dependencies

TensorFlow for macOS and Metal plug-in:

python -m pip install tensorflow==2.15.0 tensorflow-metal

Librosa and ffmpeg:

python -m pip install librosa resampy

Verify

Run the example. It will take a while the first time you run it. Subsequent runs will be faster.

python analyze.py
ℹ️
Now, you can install and use birdnetlib.

7. Usage

7.1. Usage (CLI)

  1. Inspect config file for options and settings, especially inference settings. Specify a custom species list if needed and adjust the number of threads TFLite can use to run the inference.

  2. Run analyzer.py to analyze an audio file. You need to set paths for the audio file and selection table output. Here is an example:

    python3 analyze.py --i /path/to/audio/folder --o /path/to/output/folder
    ℹ️
    Your custom species list has to be named 'species_list.txt' and the folder containing the list needs to be specified with --slist /path/to/folder. You can also specify the number of CPU threads that should be used for the analysis with --threads <Integer> (e.g., --threads 16). If you provide GPS coordinates with --lat and --lon, the custom species list argument will be ignored.

    Here’s a complete list of all command line arguments:

    --i, Path to input file or folder. If this is a file, --o needs to be a file too.
    --o, Path to output file or folder. If this is a file, --i needs to be a file too.
    --lat, Recording location latitude. Set -1 to ignore.
    --lon, Recording location longitude. Set -1 to ignore.
    --week, Week of the year when the recording was made. Values in [1, 48] (4 weeks per month). Set -1 for year-round species list.
    --slist, Path to species list file or folder. If folder is provided, species list needs to be named "species_list.txt". If lat and lon are provided, this list will be ignored.
    --sensitivity, Detection sensitivity; Higher values result in higher sensitivity. Values in [0.5, 1.5]. Defaults to 1.0.
    --min_conf, Minimum confidence threshold. Values in [0.01, 0.99]. Defaults to 0.1.
    --overlap, Overlap of prediction segments. Values in [0.0, 2.9]. Defaults to 0.0.
    --rtype, Specifies output format. Values in ['table', 'audacity', 'r', 'kaleidoscope', 'csv']. Defaults to 'table' (Raven selection table).
    --threads, Number of CPU threads.
    --batchsize, Number of samples to process at the same time. Defaults to 1.
    --locale, Locale for translated species common names. Values in ['af', 'de', 'it', ...] Defaults to 'en'.
    --sf_thresh, Minimum species occurrence frequency threshold for location filter. Values in [0.01, 0.99]. Defaults to 0.03.
    --classifier, Path to custom trained classifier. Defaults to None. If set, --lat, --lon and --locale are ignored.
    --fmin and --fmax, Minimum and maximum frequency for bandpass filter. Defaults to 0 and 15000.
    --output_file, Path to combined Raven selection table. If set and rtype is 'table', all results will be combined into this file. Defaults to None.
    --skip_existing_results, skip files that have already been analyzed. Defaults to False.

    Here are two example commands to run this BirdNET version:

    python3 analyze.py --i example/ --o example/ --slist example/ --min_conf 0.5 --threads 4
    
    python3 analyze.py --i example/ --o example/ --lat 42.5 --lon -76.45 --week 4 --sensitivity 1.0
  3. Run embeddings.py to extract feature embeddings instead of class predictions. Result file will contain timestamps and lists of float values representing the embedding for a particular 3-second segment. Embeddings can be used for clustering or similarity analysis. Here is an example:

    python3 embeddings.py --i example/ --o example/ --threads 4 --batchsize 16

    Here’s a complete list of all command line arguments:

    --i, Path to input file or folder. If this is a file, --o needs to be a file too.
    --o, Path to output file or folder. If this is a file, --i needs to be a file too.
    --overlap, Overlap of prediction segments. Values in [0.0, 2.9]. Defaults to 0.0.
    --threads, Number of CPU threads.
    --batchsize, Number of samples to process at the same time. Defaults to 1.
    --fmin and --fmax, Minimum and maximum frequency for bandpass filter. Defaults to 0 and 15000.
  4. After the analysis, run segments.py to extract short audio segments for species detections to verify results. This way, it might be easier to review results instead of loading hundreds of result files manually.

    Here’s a complete list of all command line arguments:

    --audio, Path to folder containing audio files.
    --results, Path to folder containing result files.
    --o, Output folder path for extracted segments.
    --min_conf, Minimum confidence threshold. Values in [0.01, 0.99]. Defaults to 0.1.
    --max_segments, Number of randomly extracted segments per species.
    --seg_length, Length of extracted segments in seconds. Defaults to 3.0.
    --threads, Number of CPU threads.
  5. When editing your own species_list.txt file, make sure to copy species names from the labels file of each model.

    You can find label files in the checkpoints folder, e.g., checkpoints/V2.3/BirdNET_GLOBAL_3K_V2.3_Labels.txt.

    Species names need to consist of scientific name_common name to be valid.

  6. You can generate a species list for a given location using species.py in case you need it for reference. Here is an example:

    python3 species.py --o example/species_list.txt --lat 42.5 --lon -76.45 --week 4

    Here’s a complete list of all command line arguments:

    --o, Path to output file or folder. If this is a folder, file will be named 'species_list.txt'.
    --lat, Recording location latitude.
    --lon, Recording location longitude.
    --week, Week of the year when the recording was made. Values in [1, 48] (4 weeks per month). Set -1 for year-round species list.
    --threshold, Occurrence frequency threshold. Defaults to 0.03.
    --sortby, Sort species by occurrence frequency or alphabetically. Values in ['freq', 'alpha']. Defaults to 'freq'.

    The year-round list may contain some species, that are not included in any list for a specific week. See #211 for more details.

  7. This is a very basic version of the analysis workflow, you might need to adjust it to your own needs.

  8. Please open an issue to ask for new features or to document unexpected behavior.

  9. I will keep models up to date and upload new checkpoints whenever there is an improvement in performance. I will also provide quantized and pruned model files for distribution.

7.2. Usage (Docker)

Install docker for Ubuntu:

sudo apt install docker.io

Build Docker container:

sudo docker build -t birdnet .
ℹ️
You need to run docker build again whenever you make changes to the script.

In order to pass a directory that contains your audio files to the docker file, you need to mount it inside the docker container with -v /my/path:/mount/path before you can run the container.

You can run the container for the provided example soundscapes with:

sudo docker run -v $PWD/example:/audio birdnet analyze.py --i audio --o audio --slist audio

You can adjust the directory that contains your recordings by providing an absolute path:

sudo docker run -v /path/to/your/audio/files:/audio birdnet analyze.py --i audio --o audio --slist audio

You can also mount more than one drive, e.g., if input and output folder should be different:

sudo docker run -v /path/to/your/audio/files:/input -v /path/to/your/output/folder:/output birdnet analyze.py --i input --o output --slist input

See Usage (CLI) section above for more command line arguments, all of them will work with Docker version.

ℹ️
If you like to specify a species list (which will be used as post-filter and needs to be named 'species_list.txt'), you need to put it into a folder that also has to be mounted.

7.3. Usage (Server)

You can host your own analysis service and API by launching the server.py script. This will allow you to send files to this server, store submitted files, analyze them and send detection results back to a client. This could be a local service, running on a desktop PC, or a remote server. The API can be accessed locally or remotely through a browser or Python client (or any other client implementation).

  1. Install one additional package with pip3 install bottle.

  2. Start the server with python3 server.py. You can also specify a host name or IP and port number, e.g., python3 server.py --host localhost --port 8080.

    Here’s a complete list of all command line arguments:

    --host, Host name or IP address of API endpoint server. Defaults to '0.0.0.0'.
    --port, Port of API endpoint server. Defaults to 8080.
    --spath, Path to folder where uploaded files should be stored. Defaults to '/uploads'.
    --threads, Number of CPU threads for analysis. Defaults to 4.
    --locale, Locale for translated species common names. Values in ['af', 'de', 'it', ...] Defaults to 'en'.
    ℹ️
    The server is single-threaded, so you’ll need to start multiple instances for higher throughput. This service is intented for short audio files (e.g., 1-10 seconds).
  3. Query the API with a client. You can use the provided Python client or any other client implementation. Request payload needs to be multipart/form-data with the following fields: audio for raw audio data as byte code, and meta for additional information on the audio file. Take a look at our example client implementation in the client.py script.

    This script will read an audio file, generate metadata from command line arguments and send it to the server. The server will then analyze the audio file and send back the detection results which will be stored as a JSON file.

    Here’s a complete list of all command line arguments:

    --host, Host name or IP address of API endpoint server.
    --port, Port of API endpoint server.
    --i, Path to file that should be analyzed.
    --o, Path to result file. Leave blank to store with audio file.
    --lat, Recording location latitude. Set -1 to ignore.
    --lon, Recording location longitude. Set -1 to ignore.
    --week, Week of the year when the recording was made. Values in [1, 48] (4 weeks per month). Set -1 for year-round species list.
    --overlap, Overlap of prediction segments. Values in [0.0, 2.9]. Defaults to 0.0.
    --sensitivity, Detection sensitivity; Higher values result in higher sensitivity. Values in [0.5, 1.5]. Defaults to 1.0.
    --pmode, Score pooling mode. Values in ['avg', 'max']. Defaults to 'avg'.
    --num_results, Number of results per request.
    --sf_thresh, Minimum species occurrence frequency threshold for location filter. Values in [0.01, 0.99]. Defaults to 0.03.
    --save, Define if files should be stored on server. Values in [True, False]. Defaults to False.
  4. Parse results from the server. The server will send back a JSON response with the detection results. The response also contains a msg field, indicating success or error. Results consist of a sorted list of (species, score) tuples.

    This is an example response:

    {"msg": "success", "results": [["Poecile atricapillus_Black-capped Chickadee", 0.7889], ["Spinus tristis_American Goldfinch", 0.5028], ["Junco hyemalis_Dark-eyed Junco", 0.4943], ["Baeolophus bicolor_Tufted Titmouse", 0.4345], ["Haemorhous mexicanus_House Finch", 0.2301]]}
    ℹ️
    Let us know if you have any questions, suggestions, or feature requests. Also let us know when hosting an analysis service - we would love to give it a try.

7.4. Usage (GUI)

We provide a very basic GUI which lets you launch the analysis through a web interface.

GUI screenshot
Figure 1. Web based GUI
  1. You need to install two additional packages in order to use the GUI with pip3 install pywebview gradio

  2. Launch the GUI with python3 gui.py.

  3. Set all folders and parameters, after that, click 'Analyze'.

8. Training

You can train your own custom classifier on top of BirdNET. This is useful if you want to detect species that are not included in the default species list. You can also use this to train a classifier for a specific location or season. All you need is a dataset of labeled audio files, organized in folders by species (we use folder names as labels). This also works for non-bird species, as long as you have a dataset of labeled audio files. Audio files will be resampled to 48 kHz and converted into 3-second segments (we will use the center 3-second segment if the file is longer, we will pad with random noise if the file is shorter). We recommend using at least 100 audio files per species (although training also works with less data). You can download a sample training data set here.

  1. Collect training data and organize in folders based on species names.

  2. Species labels should be in the format <scientific name>_<species common name> (e.g., Poecile atricapillus_Black-capped Chickadee), but other formats work as well.

  3. It can be helpful to include a non-event class. If you name a folder 'Noise', 'Background', 'Other' or 'Silence', it will be treated as a non-event class.

  4. Run the training script with python3 train.py --i <path to training data folder> --o <path to trained classifier model output>.

    Here is a list of all command line arguments:

    --i, Path to training data folder. Subfolder names are used as labels.
    --o, Path to trained classifier model output.
    --crop_mode, Crop mode for training data. Values in ['center', 'first', 'segments']. Defaults to 'center'.
    --crop_overlap, Overlap of training data segments in seconds if crop_mode is 'segments'. Defaults to 0.
    --epochs, Number of training epochs. Defaults to 50.
    --batch_size, Batch size. Defaults to 32.
    --val_split, Validation split ratio. Defaults to 0.2.
    --learning_rate, Learning rate. Defaults to 0.001.
    --hidden_units, Number of hidden units. Defaults to 0. If set to >0, a two-layer classifier is used.
    --dropout, Dropout rate. Defaults to 0.
    --mixup, Whether to use mixup for training.
    --upsampling_ratio, Balance train data and upsample minority classes. Values between 0 and 1. Defaults to 0.
    --upsampling_mode, Upsampling mode. Can be 'repeat', 'mean' or 'smote'. Defaults to 'repeat'.
    --model_format, Model output format. Can be 'tflite', 'raven' or 'both'. Defaults to 'tflite'.
    --model_save_mode, Model save mode. Can be 'replace' or 'append'. Defaults to 'replace'.
    --cache_mode, Cache mode. Can be 'none', 'load' or 'save'. Defaults to 'none'.
    --cache_file, Path to cache file. Defaults to 'train_cache.npz'.
    --threads, Number of CPU threads.
    --fmin and --fmax, Minimum and maximum frequency for bandpass filter. Defaults to 0 and 15000.
    --autotune, Whether to use automatic hyperparameter tuning (this will execute multiple training runs to search for optimal hyperparameters).
    --autotune_trials, Number of training runs for hyperparameter tuning. Defaults to 50.
    --autotune_executions_per_trial, The number of times a training run with a set of hyperparameters is repeated. Defaults to 1.
  5. After training, you can use the custom trained classifier with the --classifier argument of the analyze.py script. If you want to use the custom classifier in Raven, make sure to set --model_format raven.

    ℹ️
    Adjusting hyperparameters (e.g., number of hidden units, learning rate, etc.) can have a big impact on the performance of the classifier. We recommend trying different hyperparameter settings. If you want to automate this process, you can use the --autotune argument (in that case, make sure to install keras_tuner with pip3 install keras-tuner).

    Example usage (when downloading and unzipping the sample training data set):

    python3 train.py --i train_data/ --o checkpoints/custom/Custom_Classifier.tflite
    python3 analyze.py --classifier checkpoints/custom/Custom_Classifier.tflite
    ℹ️
    Setting a custom classifier will also set the new labels file. Due to these custom labels, the location filter and locale will be disabled.
  6. You can include negative samples for classes by prefixing the folder names with a '-' (e.g., -Poecile atricapillus_Black-capped Chickadee). Do this with samples that definitely do not contain the species. Negative samples will only be used for training and not for validation. Also keep in mind that negative samples will only be used when a corresponding folder with positive samples exists. Negative samples cannot be used for binary classification, instead include these samples in the non-event folder.

  7. To train with multi-label data separate the class labels with commas in the folder names (e.g., Poecile atricapillus_Black-capped Chickadee, Cardinalis cardinalis_Northern Cardinal). This can also be combined with negative samples as described above. The validation split will be performed combination of classes, so you might want to ensure sufficient data for each combination of classes. When using multi-label data the upsampling mode will be limited to 'repeat'.

9. Funding

This project is supported by Jake Holshuh (Cornell class of ’69) and The Arthur Vining Davis Foundations. Our work in the K. Lisa Yang Center for Conservation Bioacoustics is made possible by the generosity of K. Lisa Yang to advance innovative conservation technologies to inspire and inform the conservation of wildlife and habitats.

The German Federal Ministry of Education and Research is funding the development of BirdNET through the project "BirdNET+" (FKZ 01|S22072). Additionally, the German Federal Ministry of Environment, Nature Conservation and Nuclear Safety is funding the development of BirdNET through the project "DeepBirdDetect" (FKZ 67KI31040E).

10. Partners

BirdNET is a joint effort of partners from academia and industry. Without these partnerships, this project would not have been possible. Thank you!

Logos of all partners
Figure 2. Our partners