Some questions about ABM processing in the GSC module about athena-signal HOT 4 CLOSED

I guess you mean 'steering output' by each sensor channel. So yes, as you think, 'gscabm->xrefdline' should be each microphone signal, and 'gscabm->ytmp' is the filtered result from the output of FBF. We find it also works in the opposite way.
I recall that we have tested both ways, and found the results hardly make any difference in output SNR. We choose this ABM because it leads to a little lower WER in An ASR dataset.

from athena-signal.

Thanks for sharing these technical details of the algorithm, and it is interesting. I have another question, as far as I know, GSC is not very effective for non-coherence noise, so a post-filter is usually needed. However, there are no post-filter founded in the whole process. So I guess the following NS module will remove the non-coherence noise, or the post-filter will lead to lower WER? So is my guess right?

from athena-signal.

NS is an option for post-filtering, with limited effect. For ASR mode, we prefer to use a light-NS, say MCRA, to remove SOME noise with hardly speech distortion and to add noise to the training set. For communication mode, NN-based noise suppression leads to higher output SNR. It will be submitted soon.

from athena-signal.

Thanks!

from athena-signal.