Topic: Eliminating crowd noise, with artificial intelligence?

[ballerusk]

I would assume you could listen to the concert while working on other aspects of your recording and make a note where there are unwanted bits so when working in UVS5 you can just jump to those spots.

[ballerusk]

I tried using UVR5 on an old recording and found the "best" result using the following settings:

Choose process method: Demucs
Choose stems: Vocals
Segment: default
Choose demucs model: v4 htdemucs

The result wasn't perfect, but I guess it varies depending on what you feed the program. UVR5 managed to separate some chatter, but most of the "background chatter" seemed to be integrated in the instrumental-file so for me it worked best with clear speech. It also made it a little easier to "see" the talking when viewed in Izotope RX. Not sure how much I will use this going forward, unless there is very obvious talking.

Processed a 1h7m file with the settings above, it took about 1h38m (1.47x realtime) from start to finish on my 2019 iMac (3GHz 6-core i5 40GB DDR4 RAM).

[robgronotte]

It seems to be very good at removing clear talking. Very quiet talking is hit or miss, though generally an improvement.

I have found demucs and MDX give very similar results. The other choice (VR Architecture) seems to work differently, so some talk is better removed with that one, and some better with Demucs or MDX. You may want to try running a passage through both filters to get more removed.