ishine/das

Deep Audio Segmenter

Deep Audio Segmenter

DAS is a method for automatically annotating song from raw audio recordings based on a deep neural network. DAS can be used with a graphical user interface, from the terminal, or from within python scripts.

If you have questions, feedback, or find bugs please raise an issue.

Please cite DAS as:

Elsa Steinfath, Adrian Palacios, Julian Rottschäfer, Deniz Yuezak, Jan Clemens (2021). Fast and accurate annotation of acoustic signals with deep neural networks. bioRxiv, https://doi.org/10.1101/2021.03.26.436927

Installation

Pre-requisites

Anaconda: DAS is installed using an anaconda environment. For that, first install the anaconda python distribution (or miniconda).

If you have conda already installed, make sure you have conda v4.8.4+. If not, update from an older version with conda update conda.

CUDA libraries for using the GPU: While DAS works well for annotating song using the CPU, a GPU will greatly improve annotation speed and is recommended for training a DAS network. The network is implemented in the deep-learning framework Tensorflow. To make sure that Tensorflow can use your GPU, the required CUDA libraries need to be installed. See the tensorflow docs for details.

Libsoundfile on linux: The graphical user interface (GUI) reads audio data using soundfile, which relies on libsndfile. libsndfile will be automatically installed on Windows and macOS. On Linux, the library needs to be installed manually with: sudo apt-get install libsndfile1. Note that DAS will work w/o libsndfile but will not be able to load exotic audio formats.

Install DAS with or without the GUI

Create an anaconda environment called das that contains all the required packages, including the GUI:

conda env create ncb/das-env -n das

If you do not need the graphical user interface, for instance, when training DAS on a server, install the plain version:

conda env create ncb/das-nogui-env -n das

Usage

To start the graphical user interface:

conda activate das
das gui

The documentation at https://janclemenslab.org/das/ provides information on the usage of DAS:

• A quick start tutorial walks through all steps from manually annotating song, over training a network, to generating new annotations.
• How to use the graphical user interface.
• How to use DAS from the terminal or from python scripts.

Acknowledgements

The following packages were modified and integrated into das:

• Keras implementation of TCN models modified from keras-tcn (in das.tcn)
• Trainable STFT layer implementation modified from kapre (in das.kapre)

See the sub-module directories for the original READMEs.