Expectation-maximization algorithms for IS-NMF

This repository contains the code related to our paper titled Expectation-maximization algorithms for Itakura-Saito nonnegative matrix factorization published at Interpseech 2018. If you use the content of this repository, please cite our paper.

Setup

This code was initially developed with Matlab, but we have adapted it to Octave. To use it, you need to install some Octave packages as follows:

sudo apt install liboctave-dev
sudo apt install octave-control
sudo apt install octave-signal
sudo apt install octave-dataframe	

Note that you can ignore dataframe if you don't intend to display the results using this tool.

The experiments use data from the GRID corpus. This dataset is devoted to audio-visual speech transcription, but here we only use the audio data, thus one only needs to download the audio at 25 kHz for a given pair of speakers, and place the unziped files in the data/grid/ folder. In our paper, we used speakers s1 and s4. You can use a different folder structure and/or speaker(s) as long as you change the path and speaker indices accordingly in the global_setup.m file.

Usage

To reproduce the results from our paper, run the scripts in order:

• 1.prepare_data.m creates the mixtures and the dictionary audio files.
• 2.dico_learning.m computes the NMF dictionaries for several dic. sizes.
• 3.separation.m performs supervised source separation using the various NMF algorithms, and computes the SDR/SIR/SAR scores.
• 4.display_results.m produces the figures and table from the paper.