Iterative combined LCMV-TRINICON beamforming for separating multiple speech sources in noisy and reverberant environments

4 speakers example [Scenario 1]

The LCMV-TRINICON [LST] is compared with:

[MCNMF] A. Ozerov and C. Fevotte, “Multichannel nonnegative matrix factorization in convolutive mixtures for audio source separation,” IEEE Trans. Audio, Speech and Language Processing, vol. 18, no. 3, pp. 550–563, Mar. 2010.

Available: http://www.irisa.fr/metiss/ozerov/Software/multi_nmf_toolbox.zip.

We would like to express our gratitude to D. Kounades-Bastian for sharing his code, extending the original MCNMF implementation by A. Ozerov and C. Fevotte from stereophonic input to any number of microphones.

Two types of scenarios are considered for the testing:

Filtering source signals with impulse responses (IRs) measured at the acoustic lab at Bar-Ilan University (taken from the BIU database)
Real-life recordings of real persons in the same acoustic lab.

2 speakers example [Scenario 1]:

In - Mic. 1 1:03 MCNMF - out 1 1:03 MCNMF - out 2 1:03 LST - out 1 1:03 LST - out 2 1:03

4 speakers example [Scenario 1]:

In - mic1 1:03 MCNMF - out 1 1:03 MCNMF - out 2 1:03 MCNMF - out 2 1:03 MCNMF - out 4 1:03 LST - out 1 1:03 LST - out 2 1:03 LST - out 3 1:03 LST - out 4 1:03

3 speakers example [Scenario 2]:

In - mic 1 1:19 MCNMF - out 1 1:19 MCNMF - out 2 1:19 MCNMF - out 3 1:19 LST - out 1 1:19 LST - out 2 1:19 LST - out 3 1:19