Topic: A.I. audio stem software

[AbbyTaper]

Hope they have a new year sale or something on Spectralayers. That unmix crowd feature seems golden for audience mumbling chatter.

I tried it once, and would describe the results as "mixed".

[robgronotte]

Hope they have a new year sale or something on Spectralayers. That unmix crowd feature seems golden for audience mumbling chatter.

I tried it once, and would describe the results as "mixed".

I have found the whole program very difficult to use. Very disappointing, because the features seem great.

[lavaux]

Based on quick research, seems like SpectralLayers was built off of Spleeter, which is the old method of extracting elements. They now use Demucs, which you get for free with UVR (but UVR doesn't have the vocal learning feature that Spectralayers has.

I think MVSEP and X-Minus are the most cutting edge, with UVR right behind them. I use UVR to remove audience, then load it up into Izotope RX and spectrally remove all the talking/coughing/non-applause.

Here's a really great research document that showcases all the current updates and movements in the AI audio separation field:
https://docs.google.com/document/d/17fjNvJzj8ZGSer7c7OFe_CNfUKbAxEh_OBv94ZdRG5c/edit?usp=sharing

[if_then_else]

There are a few other options that haven't been mentioned in this thread yet.

• Remix from Acon Digital (https://acondigital.com/products/remix) that's also part of Acoustica Premium Edition. I've got the latter and the tool works great for balancing instruments in a mix. But it wouldn't be my top choice for stem separation.
  
• Rip X (https://hitnmix.com) which - according to some reviews by sound and mixing engineers - is supposedly the most advanced tool of them all. That said, the DAW itself in which stem separation is integrated looks like it could need improving.
  

[phil_er_up]

Have used RX, UVR and SL using many different files and genres. Used the Demus in UVR and found that SL and UVR sounded the best in stem separation and almost similar. RX sometimes has this metallic sound to it. Though found mixed results depending on the genre I used with each software. With RX and UVR they just split stems and that is it. No editing features. SL has these editing features like RX with the addition of the eraser tool, batch processing, select similar and ability to add layers or extract from layers and add to another layer or create a new layer. This is a big step IMO.

Have a fast PC - time to process 70 minute set in following software with 4 stem extraction:
RX - 8 minutes
UVR - 8 minutes
SL 60-70 minutes - you can get around the slow processing time by using the batch processing tool and run it over night.

Tried this over the weekend took a audience recording with crowd noise in the back ground and ran "unmix crowd noise" in SL. Was surprised  how well it took out the chatter from the floor noise. What was left after extracting the crowd noise almost sounded like a SBD though some vocals were taken out with the crowd noise. Then selected the vocals manually in a short section in the crowd noise layer and  ran "select similar". It ran for hours to try and find all of them. Did an ok job. If you extracted just the vocal then ran "unmix crowd noise" it might make a nice sounding audio file. SL is like photoshop and can do many more features then just splitting the files. Though as robgronotte suggests its not as easy program to use and would take time to figure out how to really use it.

====================================================

Thanks to  lavaux and  if_then_else for their input and comments. It is appreciated.

Will look into the software you guys posted about. Thank you!

[robgronotte]

Well, I still haven't figured out a lot of what can be done with SpectraLayers 11, but I just tried using the "unmix crowd noise" for the first time with excellent results.

I took a 14 second part of a recording where there was a lot of crowd chatter under the main vocals and ran it through UVR5.  Took the vocal file from there and ran the unmix crowd noise in SL11, and it gave me a pretty much perfect separation between the singer and the background chatter.  It also was very quick doing that, although again it was only 14 seconds.   Then I exported the "Foreground" layer, and mixed it back with the instrumental audio file from UVR5.

I'm hoping I can use this method in the future instead of agonizingly trying to remove chatter from vocal sections using the spectral repair feature of iZotope RX.

[phil_er_up]

Well, I still haven't figured out a lot of what can be done with SpectraLayers 11, but I just tried using the "unmix crowd noise" for the first time with excellent results.

I took a 14 second part of a recording where there was a lot of crowd chatter under the main vocals and ran it through UVR5.  Took the vocal file from there and ran the unmix crowd noise in SL11, and it gave me a pretty much perfect separation between the singer and the background chatter.  It also was very quick doing that, although again it was only 14 seconds.   Then I exported the "Foreground" layer, and mixed it back with the instrumental audio file from UVR5.

I'm hoping I can use this method in the future instead of agonizingly trying to remove chatter from vocal sections using the spectral repair feature of iZotope RX.

Good job figuring out how to use the tools. Cool that you used SL then UVR5 then SL.

That is my thought pattern too. Will have to use the tools together to get what you want. RX and UVR5 are good at separating a file into stems though that is all they do. SL does so much more though you have to use their tools then manually edit the output to separate what you want in the file and (Noise, chatter, instruments) that you want to take out.

Tried "unmix crowd noise" and thought it did a good job. It completely took out chatter between songs. Though in some songs it took out some of the vocals that the program thought was crowd noise. Wish there was a way to only select different sections of the file then run the tool.