Christoph Boeddeker

No, it is not necessary to know the noise. You can use a heuristic to identify the noise or you can reduce the number of speakers to the actual number...

Since you don't give me any context, it is probably the best for you, to reduce `nsources` to the actual number of speakers. This will probably degenerate the performance, but...

I personally prefer to use `nsources=4` in this case. But since it looks like you have an application in mine, where you don't have the source signals (i.e. no academic...

> yep that's fine. So to recap, for nsources==3, it's ok not to have the noise Yes. If the SNR is ok and the noise is not a point source....

Sorry, I forgot, that we have two versions online. In version in `nara_wpe` is slightly older. The function `paderbox.transform.module_stft.istft` does the same, but recently we added the `num_samples` argument. Nevertheless,...

I would say, try to keep it simple and "low level" (i.e. one equation, one function). With `apply_beamforming` you get, as popcornell said, most users and maybe there is no...

Hi, difficult to say, what went wrong. I need more information to get an idea. Have you tried offline WPE. What ist the difference of your data, compared to the...

This difference sounds to be too large. Nevertheless, there are some reasons, why numpy is/can be faster: - The tensorflow code is similar to on an older numpy implementation, that...

> My tensorflow version is 1.13, did you mean the conjugate computation while calculating R is ignored? This can happen. The problem is that tensorflow does some optimizations of the...

I published that torch code, because the pytorch people wanted to have an example, that uses complex numbers. In #46 I accidentally merged the code and forgot it. I checked...