Seperation of Multiple Sources. about pyroomacoustics

[fakufaku]

Hi Philip,
Awesome! I really appreciate the feedback on the package, so please keep it coming.
As to FRIDA, the algorithm is a little more finicky than SRP or MUSIC to work well. In our case, we had to restrict the frequency range to higher frequencies, typically 2 to 5 kHz, to get good performance. This might vary depending on the geometry of your array.
We have the files we used to run the simulation in the paper on github too, so you can check the configuration of the algorithm there.
https://github.com/LCAV/FRIDA/blob/master/figure_doa_separation.py#L150
Please let me know if you manage to make it work, and like-wise if it doesn't :)

[jay-pee]

Thank you I will have a look into it and will keep you updated :)

In the moment I don't fully understand FRIDA and TOPS algorithm. Have to read the linked paper more intensively...

But is there another way besides try and error to get the right configuration. In example for SRP and Music you can have a look at the steered power spectrum to see which bins may be suited (See plot). Is there a method to analyse which bins are right?

In the plot you can also see another beamformer which is the SRP method with a Capon beamformer. I will push it to the repository when I fixed a few bugs and found a faster implementation.

[fakufaku]

Nice plots! And looking forward to the Capon implementation.

I don't have a systematic way of doing it. In general, you have to match the wavelength of the signal to the distances between the microphones in your array. If the wavelength is very long compared to the inter-mic separation, then there is barely any phase difference between microphones and localization will be poor, as you can observe on the bottom parts of the graphs. On the contrary, if the wavelength is much shorter than the distance, you start to get aliasing, i.e. there is a phase ambiguity. Your target is the sweet spot between the two regions.

[fakufaku]

Hi @jay-pee , did you manage to have FRIDA work ?

[jay-pee]

Hey @fakufaku sorry for my late answer. My master thesis was more about getting the full spatial spectrum over time and classify this data and do a tracking over time, so you can separate sources over time. Basically, it as a post processing of the data you get out of different DOA estimation methods. I haven't tried much but if FRIDA is aiming more on the sweet spot between low resolution and aliasing it wasn't the right DOA estimation method for me because I wanted the broadband spectrum of DOA estimations. In September, I will start as a PHD student and I hope I can contribute more then I did as a Master student. Capon is still on my list for implementation.
I think you can close this issue if you want.

[fakufaku]

@jay-pee Thanks for the feedback. It seems indeed that this might not be a good fit for FRIDA. Hope to hear from you in the future then! All the best with the start of your PhD.