Please take a look at our latest BirdNET repository with updated models, tools and interfaces: BirdNET-Analyzer
TFLite version of BirdNET. Bird sound recognition for more than 6,000 species worldwide.
This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.
K. Lisa Yang Center for Conservation Bioacoustics, Cornell Lab of Ornithology, Cornell University
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
TFLite for x86 platforms comes with the standard Tensorflow package. If you are on a different platform, you need to install a dedicated version of TFLite (e.g., a pre-compiled version for Raspberry Pi).
We need to setup TF2.3+ for BirdNET. First, we install Python 3 and pip:
sudo apt-get update
sudo apt-get install python3-dev python3-pip
sudo pip3 install --upgrade pip
Then, we can install Tensorflow with:
sudo pip3 install tensorflow
TFLite on x86 platform currently only supports CPUs.
Note: Make sure to set CUDA_VISIBLE_DEVICES="" in your environment variables. Or set os.environ['CUDA_VISIBLE_DEVICES'] = '' at the top of your Python script.
In this example, we use Librosa to open audio files. Install Librosa with:
sudo pip3 install librosa
sudo apt-get install ffmpeg
You can use any other audio lib if you like, or pass raw audio signals to the model.
If you don't use Librosa, make sure to install NumPy:
sudo pip3 install numpy
Note: BirdNET expects 3-second chunks of raw audio data, sampled at 48 kHz.
You can run BirdNET via the command line. You can add a few parameters that affect the output.
The input parameters include:
--i, Path to input folder. All the nested folders will also be processed.
--o, Path to output folder. By default results are written into the input folder.
--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 to ignore.
--overlap, Overlap in seconds between extracted spectrograms. Values in [0.0, 2.9]. Defaults tp 0.0.
--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.
--custom_list, Path to text file containing a list of species. Not used if not provided.
--filetype, Filetype of soundscape recordings. Defaults to 'wav'.
--num_predictions, Defines maximum number of written predictions in a given 3s segment. Defaults to 10
Note: A custom species list needs to contain one species label per line. Take a look at the model/label.txt for the correct species label. Only labels from this text file are valid. You can find an example of a valid custom list in the 'example' folder.
Here are two example commands to run this BirdNET version:
python3 analyze.py --i 'example/XC558716 - Soundscape.mp3' --lat 35.4244 --lon -120.7463 --week 18
python3 analyze.py --i 'example/XC563936 - Soundscape.mp3' --lat 47.6766 --lon -122.294 --week 11 --overlap 1.5 --min_conf 0.25 --sensitivity 1.25 --custom_list 'example/custom_species_list.txt'
Note: Please make sure to provide lat, lon, and week. BirdNET will work without these values, but the results might be less reliable.
The results of the anlysis will be stored in a result file in CSV format. All confidence values are raw prediction scores and should be post-processed to eliminate occasional false-positive results.
Please don't hesitate to contact us if you have any issues with the code or if you have any other remarks or questions.
Our e-mail address: ccb-birdnet@cornell.edu
We are always open for a collaboration with you.
This project is supported by Jake Holshuh (Cornell class of ’69). The Arthur Vining Davis Foundations also kindly support our efforts.