New Music Source Separation Website

[zeerround]

https://www.lalal.ai/

Pros:

Better sounding stems (at least on my test track) than Spleeter Gui (Vocal, Bass) or Demucs (Drums)
Also does Acoustic and Electric Guitar stems!
You don't have to mess with python, command line, etc.

Cons:

You can only get one stem type at a time, and then have to make your own "other", with the original and all your stems (easy enough, depending on your audio toolset).
Pay by the minute of source material (times the number of stems you want).

$10 for 90 minutes of source or $20 for 300 min.​

For my test track I made these stems:

Vocal​

Acoustic Guitar​

Electric Guitar​

Piano (although buried in the mix in my test track)​

Bass​

Drums​

Then an "Other" stem, mixing all the above and subtracting from the original track (or mixing with inverse phase if you prefer).

I used all those stems, and SpecWeb, to make a 7.1.4 Dolby Atmos up-remix.

(If you have a Dolby Atmos rig) PM me for a sample.

Oh, one other setting they have is easy to miss; Mild, Normal, and Aggressive filtering. Normal is the default. The English description of what that does is not very clear, but I think it works like a noise gate after the separation. e.g. aggressive would cut out more "stuff" that in theory you don't want, but also might be stuff you do. For my test I ended up using Normal for everything except vocals, where I used mild. More aggressive settings on the vocals had some missing syllables, which made the sound jump in the up-remix. I also hand edited out some noise/bleed from the mild vocal stem, before creating my "other" stem.

 

[winopener]

Several question:
1) Processing speed?
2) resolution (96/24, 16/48....)
3) input/output format?

 

[zeerround]

1) Processing speed? - In my experience it took about 5min per stem, for a 6:46min track. Most of that time is waiting in the que.
2) resolution (96/24, 16/48....) - I only tested 44.1/16 (most/all of these tools downsample to 44.1(?))
3) input/output format? - They list them. I tested wav and flac.

 

[winopener]

Thanks, very appreciated.

 

[zeerround]

I'm experimenting with a different approach to remixing the separated stem outputs from LALAL.AI.

Instead of creating an "Other Stem" (all the stems mixed together and then subtracted from the original stereo) I take advantage of the "stem minus" output of LALAL.AI. e.g. For each separation, say Piano, you also get the original stereo with no Piano.

Now here's the new idea; mix each stem with it's "stem minus", and use that for up/remix as a blended source. The level of the "stem minus" should be such that all of the stems you have would add to unity gain. e.g. if you have 5 stems, 1/5 = .2 or -14dB. You mix each stem with its "minus stem" -14dB down.

If you have 4 stems, mix each (0dB) with its "minus stem" at -12dB
For 5 stems, mix each (0dB) with its "minus stem" at -14dB
For 6 stems, mix each (0dB) with its "minus stem" at -15.6dB

Having the "minus stem" blended in with each separated source helps reduce any objectionable separation artifacts, and doesn't seem to to detract too much from separation, in my tests so far. Any sounds that were not in the separations add back up to their original volume from the blends.

Also not having to worry about where to place an "Other" stem is nice, when working in 7.1.4 or higher.

I found a test track that I could get 6 good stems from: Porcupine Tree - Lazarus (Live); having Piano, Acoustic and Electric Guitars, Bass, Drums, and Vocals.

I made some upmixes from the blended stems:

Electric Guitar -> Quad
Drums -> Quad
Piano -> Quad
Vocal -> 5.1 (in this case pretty much all center)

and left the acoustic guitar and bass stems in stereo.

And arranged things (7.1.4):

Acoustic Guitar Stereo Rears
Bass Stereo crossovers Sides
Electric Guitar Quad Heights
Drums Quad Heights
Piano Quad Fronts and Rear Heights
Vocals 5.1 with Rears in Heights

I did have to manually edit a stem in audition, to stabilize the sound field for the guitar solo.

LALAL.AI has removed some options from their web interface, so you no longer have any knobs to turn (but maybe still available if you use their API instead?). I'll have to dig into that to see if a less aggressive filter setting is still available and if that would help remove the need to edit stems.

So far so good.

Using AudioMuxer and AWS media converter I can then make a 7.1.4 Atmos mp4 file of the same format and quality as you get from Tidal Atmos.

Yes I know this album is available in 5.1, but I'm working on Stereo to 7.1.4 techniques so...