Audioshake

 

Paid, $16 per wav stem, 2 or 5 stems (6? (guitar and piano) or 4 stems for preview (Indie creators)

 

Better piano model than GSEP.

"gsep piano model is very clean but sometimes fails in bigger mix, when there are a lot of instruments"

And also guitar stem

 

 

Instead of Audioshake you can use:

- myxt.com (also paid, 3 stem model, prob. 16kHz cutoff which Audioshake normally doesn't have. No other stem. Results, maybe slightly better than Demucs)

- Algoriddim djay pro

-  Neural Mix Pro (part of Algoriddim, also uses Audioshake), but it’s only for MacOS
- LANDR Stems (cheaper, also uses Audioshake; plugin, probably doesn’t work locally, free access won’t give you access to stems; “LANDR Stems is only included in Studio Standard + Studio Pro” it’s not included in trial; SDR: 1 | 2)

- https://twoshot.app/model/289

 

Audioshake is suspected that it is just MDX with expanded dataset, but there’s no evidence at the moment. Comparing to UVR/MDX-UVR NET 1 model, vocal stem is 9.793 vs 9.702 in free MDX-UVR, so they’re close as for vocals.

Their researcher said they were training UMXHQ model at this period of time of 2020 Demixing Challenge.

 

Free Demucs 3 has a much better SDR for drums and bass than Audioshake, however the SDR for vocals and others is worse.

 

It accepts only non-copyrighted music for separations, but you can slow it down to circumvent it (some music like K-Pop BTS is not detected) but changing speed to 110% yields better results, even vs reversing the track.

 

Upload limit is one minute, so theoretically you can cut and merge chunks, but AS will fade out each chunk, so you need to find specific overhead to begin every next chunk with, to merge chunks seamlessly (I don’t remember if it solves the problem of AS watermark, though).

Then, you can download preview (chunk) for free using similar method like described in allflac section (Chrome Dev Tools -> Network -> Set filter to amazon) but result file is unfortunately only 128kbps mp3.

 

They are now limiting how many audio files you can upload to preview, but that can easily be mitigated by just using a temporary email provider or adding “+1” or “+2” or “.” to your gmail address, so you will still receive your email e.g. y.o.u.r.m.a.i.l.@gmail.com is the same for Google as yourmail@gmail.com.

 

You can also ping Smudge, Baul Plart or Bas Curtiz in #request-separation to isolate some song to make this all easily just for you (edit. 09.02.2023 - at least the Bas’ tool stopped working, so the rest like AS Tool might be dead too - at least in terms of API access, not sure).