Download the system

Download the latest package or clone the system directly from the repository. To clone repository with HTTPS:

git clone

or with ssh:

git clone

The system is developed for Python 2.7, Python 3.5 and Python 3.6. The system is tested to work on Linux, Windows and MacOS platforms.

One can install either the official CPython or use some Python distribution based on it. New users are recommended to use Anaconda Python distribution.

Using the system on Windows

The baseline system uses rather long directory paths, since it is storing 32 character MD5 hash of the system parameters into directory names. Some Windows system have path length limit (lowest 260 characters), which is causing problems. To avoid problems related to this, install the system as close as possible to the drive root.

External modules

To ensure that all external modules are installed, run command:

pip install -r requirements.txt


PySoundFile is used to read and write audio files in the system. The library depends on system level library libsndfile. Under Linux (Debian/Ubuntu) you can install this with command:

sudo apt-get install libsndfile1


The system uses by default Theano as Keras backend.

There was a bug in Theano 0.8.2 version, so make sure to use 0.9.0 release. The bug will affect Task 2 system (when using GPU and binary_crossentropy as loss function). To fix this, make sure you have installed correct version of Theano:

pip install theano==0.9.0

Or use latest from git with command:

pip install --upgrade --no-deps git+git://


It is also possible to use Tensorflow as Keras backend. If you plan to use it with GPU make sure you have installed GPU enabled package:

pip uninstall tensorflow
pip uninstall tensorflow-gpu
pip install tensorflow-gpu

Current version of Tensorflow only supports CUDA 3.0 compatible graphic cards.

Audio datasets

The system will automatically download the needed audio datasets, and place them under the directory specified in the parameters (see parameter path->data).

Dataset Type Audio files Size on disk License
TUT Acoustic scenes 2017, development Acoustic scene 4680 22Gb Academic use only (see EULA inside the package)
TUT Rare sound events 2017, development Sound events / synthetic 1281 9.2Gb Academic use only (see EULA inside the package)
TUT Sound events 2017, development Sound events realistic 24 2.6Gb Academic use only (see EULA inside the package)