Now published in eLife: https://elifesciences.org/articles/63853
Code to reproduce results from the article is in the directory ./article
A neural network architecture (shown below) that automates annotation of birdsong and other vocalizations by segmenting spectrograms, and then labeling those segments.
This is an example of the kind of annotations that tweetynet
learns to predict:
Short version (for details see below):
$ pip install tweetynet
$ conda install tweetynet -c conda-forge
On Windows, you need to add an additional channel, pytorch
.
You can do this by repeating the -c
option more than once.
$ conda install tweetynet -c conda-forge -c pytorch
$ # ^ notice additional channel!
Long version:
To facilitate training tweetynet
models and using trained models
to predict annotation on new datasets,
we developed the vak
library,
that is installed automatically with tweetynet
.
If you need more information about installation, please see the vak
documentation:
https://vak.readthedocs.io/en/latest/get_started/installation.html
For a tutorial on using tweetynet
with vak
, please see the vak
documentation:
https://vak.readthedocs.io/en/latest/get_started/autoannotate.html
In the directory ./article
we provide code to reproduce the results in the article
"TweetyNet: A neural network that enables high-throughput, automated annotation of birdsong"
https://elifesciences.org/articles/63853
Please see the README in that directory for instructions on how to install and work with that code.
To train models, you must supply training data in the form of audio files or
spectrogram files, and annotations.
The package can generate spectrograms from .wav
or .cbin
audio files.
It can also accept spectrograms in the form of Matlab .mat
files or .npz
files created by numpy
.
vak
uses a separate library to parse annotations, crowsetta
,
which handles some common formats and can also be used to write custom parsers for other formats.
Please see the crowsetta
documentation for more detail:
https://crowsetta.readthedocs.io/en/latest/#
It is possible to train on any manually annotated data but there are some useful guidelines:
- Use as many examples as possible - The results will just be better. Specifically, this code will not label correctly syllables it did not encounter while training and will most probably generalize to the nearest sample or ignore the syllable.
- Use noise examples - This will make the code very good in ignoring noise.
- Examples of syllables on noise are important - It is a good practice to start with clean recordings. The code will not perform miracles and is most likely to fail if the audio is too corrupt or masked by noise. Still, training with examples of syllables on the background of cage noises will be beneficial.
For more details, please see the vak documentation.
If you run into problems, please use the issue tracker or contact the authors via email in the paper above.
If you use or adapt this code, please cite its DOI:
Released under BSD license.
Thanks goes to these wonderful people (emoji key):
yardencsGitHub π» π π£ π π€ π¬ π§ |
David Nicholson π» π π£ π π€ π¬ π§ |
Zhehao Cheng π |
This project follows the all-contributors specification. Contributions of any kind welcome!