Functions for working with this data repository:

pip install birdsong-recognition-dataset==0.3.2


Build Status DOI PyPI version License


Python utility for working with data from the following repository:
Koumura, T. (2016). BirdsongRecognition (Version 1). figshare.

The repository contains .wav files of Bengalese finch song from ten birds and annotation for the songs in .xml files.

This repository provides a great resource, and was used to benchmark a sliding window-based neural network for segmenting and labeling the elements of birdsong, as described in the following paper:
Koumura, Takuya, and Kazuo Okanoya.
"Automatic recognition of element classes and boundaries in the birdsong with variable sequences."
PloS one 11.7 (2016): e0159188.

The code for the network can be found here:

The original code was released under the GNU license:

The birdsongrec module is used with the crowsetta package to make the repository a dataset available in the hybrid-vocal-classifier and vak libraries.


with pip

$ pip install birdsong-recognition-dataset

with conda

$ conda install birdsong-recognition-dataset -c conda-forge


The main thing that birdsongrec gives you is easy access to the annotation, without having to deal with the .xml file format.

This format is schematized in this XML schema file, adapted from the original under the GNU license (file is unchanged except for formatting for readability).

To access the annotation in the Annotation.xml files for each bird, use the parse_xml function.

>>> from birdsongrec import parse_xml
>>> seq_list = parse_xml(xml_file='./Bird0/Annotation.xml', concat_seqs_into_songs=False)
>>> seq_list[0]
Sequence from 0.wav with position 32000 and length 43168
>>> seq_list[0].syls[:3]
[Syllable labeled 0 at position 2240 with length 2688, Syllable labeled 0 at position 8256 with length 2784, Syllable labeled 0 at position 14944 with length 2816]  

Notice that this package preserves the abstraction of the original code, where syllables and sequences of syllables are represented as objects. This can be helpful if you are trying to replicate functionality from that code.
Importantly, each song is broken up into a number of "sequences". You can set the flag concat_seqs_into_songs to True if you want parse_xml to concatenate sequences by song (.wav file), so that each Sequence is actually all the sequences from one song.
If you are using the annotation to work with the dataset for some other purpose, you may find it more convenient to work with some other format. For that, please check out the crowsetta tool, that helps with building datasets of annotated vocalizations in a way that's annotation-format agnostic.

The birdsongrec package also provides a convenience function to load the annotation for an individual song, load_song_annot. This is basically a wrapper around parse_xml that filters out the songs you don't want.

>>> from birdsongrec import load_song_annot
>>> wav1 = load_song_annot(wav_file='1.wav')
>>> print(wav1)                                                                                                  
Sequence from 1.wav with position 32000 and length 214176  

Getting Help

Please feel free to raise an issue here:


BSD License.


If you use this package, please cite the DOI: