BirdVoxDetect: detection and classification of flight calls

BirdVoxDetect is a pre-trained deep learning system which detects flight calls from songbirds in audio recordings, and retrieves the corresponding species. It relies on per-channel energy normalization (PCEN) and context-adaptive convolutional neural networks (CA-CNN) for improved robustness to background noise. It is made available both as a Python library and as a command-line tool for Windows, OS X, and GNU/Linux.

Installation

The simplest way to install BirdVoxDetect is by using the pip package management system, which will also install the additional required dependencies if needed.

pip install birdvoxdetect

Note that birdvoxdetect requires:

• Python (3.6, 3.7, or 3.8)
• librosa (==0.7.0)
• tensorflow (>=2.2)
• scikit-learn (==0.21.2)
• birdvoxclassify (>=0.3)
• h5py
• pandas

Usage

From the command line

To analyze one file:

birdvoxdetect path/to/file.wav

To analyze multiple files:

birdvoxdetect path/to/file1.wav path/to/file2.wav

To analyze one folder:

birdvoxdetect path/to/folder

On Windows:

birdvoxdetect path\to\folder

Optional arguments:

--clip-duration CLIP_DURATION, -d CLIP_DURATION
                      Duration of the exported clips, expressed in seconds
                      (fps). The default value is 1.0, that is, one second.
                      We recommend values of 0.5 or above.
--export-clips, -c    Export detected events as audio clips in WAV format.
--export-confidence, -C
                      Export the time series of model confidence values of
                      events in HDF5 format.
--export-faults, -f   Export list of sensor faults in CSV format.
--export-logger, -l   Export output of Python logger in TXT format.
--output-dir OUTPUT_DIR, -o OUTPUT_DIR
                      Directory to save the output file(s); The default
                      value is the same directory as the input file(s).
--predict-proba, -p   Export output probabilities in JSON format.
--quiet, -q           Print less messages on screen.
--suffix SUFFIX, -s SUFFIX
                      String to append to the output filenames.The default
                      value is the empty string.
--threshold THRESHOLD, -t THRESHOLD
                      Detection threshold, between 10 and 90. The default
                      value is 50. Greater values lead to higher precision
                      at the expense of a lower recall.
--verbose, -v         Print timestamps of detected events.
--version, -V         Print version number.

From Python

Call syntax:

import birdvoxdetect as bvd    
df = bvd.process_file('path/to/file.wav')

df is a Pandas DataFrame with three columns: time, detection confidence, and species.

Below is a typical output from the test suite (file path tests/data/audio/fd79e55d-d3a3-4083-aba1-4f00b545c3d6.wav):

Time (hh:mm:ss),Detection confidence (%),Order,Order confidence (%),Family,Family confidence (%),Species (English name),Species (scientific name),Species (4-letter code),Species confidence (%)
0,00:00:08.78,70.15%,Passeriformes,100.00%,Turdidae,100.00%,Swainson's thrush,Catharus ustulatus,SWTH,99.28%

Contact

Official website

Please visit our website for more information on the BirdVox project: https://wp.nyu.edu/birdvox

The main developer of BirdVoxDetect is Vincent Lostanlen, scientist at CNRS, the French national center for scientific research.

Discussion group

For any questions or announcements related to BirdVoxDetect, please refer to our discussion group: https://groups.google.com/g/birdvox

References

Please cite the following paper when using BirdVoxDetect in your work:

Robust Sound Event Detection in Bioacoustic Sensor Networks
Vincent Lostanlen, Justin Salamon, Andrew Farnsworth, Steve Kelling, and Juan Pablo Bello
PLoS ONE 14(10): e0214168, 2019. DOI: https://doi.org/10.1371/journal.pone.0214168

As of v0.4, species classification in BirdVoxDetect relies on a taxonomical neural network (TaxoNet), which is distributed as part of the BirdVoxClassify package. For more details on TaxoNet, please refer to:

Chirping up the Right Tree: Incorporating Biological Taxonomies into Deep Bioacoustic Classifiers
Jason Cramer, Vincent Lostanlen, Andrew Farnsworth, Justin Salamon, and Juan Pablo Bello
In IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), Barcelona, Spain, May 2020.