FASST Documentation

This is a copy of FASST toolbox modified to build with QT5 on Mac OS X

License

This software is released under the Q Public License Version 1.0. Please find a copy of the license at http://opensource.org/licenses/QPL-1.0.

Install dependencies

FASST has the following dependencies:

CMake to build
Qt for the XML I/Os and the future GUI
libsndfile for the WAV file I/Os
Eigen for the linear algebra and the FFT
(optional) numpy for python scripts

On OS X

This guide assumes that you use Apple's developer tools and the brew package manager (which is trivial to install and simplifies a lot of things for developers on OS X).

Install all the dependencies with the following commands:

brew install cmake qt libsndfile eigen python
pip install numpy

Build

mkdir build
cd build
cmake ..
make

Run the examples

Example scripts will be located in the build/examples directory.

The typical workflow is to use some scripts to initialize the source model, write it to an XML file and call FASST binaries.

Run the first example

The first example script is provided in Python and MATLAB. It separates three musical sources from an instantaneous mixture.

With Python

python example.py

With MATLAB

To run the MATLAB example, simply cd to the example directory and type example

Run the second example

The second example script is provided in MATLAB. It performs source separation on noisy speech data from the first track of the 2nd CHiME challenge (http://spandh.dcs.shef.ac.uk/chime_challenge/). To run this script, you will need:

The development and/or the test dataset downloadable at http://spandh.dcs.shef.ac.uk/chime_challenge/chime2_task1.html
The following archive containing additional annotation files and models: tar and zip

In addition, you need to edit the script to define the directories where the data and the models are.

The script takes 3 arguments:

The dataset ('devel' or 'test')
The SNR level ('m6dB', 'm3dB', '0dB', '3dB', '6dB', '9dB')
The filename (something like 's1_bgaa9a.wav')

Write your own script {#write-your-own-script}

As a prerequisite, you need to be familiar with the FASST source model. If not, you can refer to this paper where it is fully described.

You have to be aware that some features that are present in the MATLAB version of FASST are still not implemented, here is a quick list:

ERB transform
GMM and HMM spectral models
Partially-adaptative spectral parameters
Only one additive noise model

FASST is decomposed in 3 different executables. Here are their descriptions and usages:

Compute mixture covariance matrix

Usage:  comp-rx input-wav-file input-xml-file output-bin-file

Estimate source parameters

Usage:  model-estimation input-xml-file input-bin-file output-xml-file

Separate sources

Usage:  source-estimation input-wav-file input-xml-file output-wav-dir

audiofilter/fasst