Looking for a BSS algorithm that works with Real World audio recordings

I am looking for a Blind Audio Source Separation algorithm that I can use to reconstruct human voices from real world recordings. My application is simultaneously recording 2 people speaking using 2 microphones and saving the recordings as mono audio files.

I have tried several implementations of FastICA (including FastICA Package for MATLAB from Aalto University, scikit learn BSS using FastICA, and several other university professors’ implementations) and a couple DUET implementations, but none have worked to separate real world recordings of 2 speakers. I know that FastICA fails when there is delay between the sources, which would be true in the real world so I am not surprised that FastICA did not work for me. Also, DUET makes the assumption that the sources are in separate frequency bands, which is not true of human voices.

Several of the algorithms have been able to separate the “manual mixing” case (speakers are recorded separately and then linearly mixed together using MATLAB). However my project requires that the voices be recorded simultaneously in real time.

I have come across others who seem to be looking for similar algorithms (Real World Blind Source Separation), and I am aware that there are other algorithms (Jade, AMUSE). My understanding is that JADE is another ICA algorithm so I expect that I would get results similar to the results I got from FastICA.

I was wondering if anyone had a BSS algorithm that has worked for reconstructing real world audio samples and could post a link to it. My project is running on a Unix-based OS that does not have MATLAB so preferably the algorithm is not coded for MATLAB, but I have access to MATLAB on another computer so I could still work with it.