https://www.splitter.ai/

How does this work?

Isolating instruments from music is now possible using AI, and Splitter is based on Deezer's open source research project Spleeter to accomplish this.

Currently a few different models are available; 2 stem, 4 stem and 5 stem models.

The 5 stem model for example, can extract vocals, drums, piano, bass and other (guitar, synths, etc...), while the 2 stem model extracts the instrumental and the vocals.

In the future we'll hopefully see more isolations.

What can I use this technology for?

This platform is intended to be just another tool in your toolbox.

Be careful when isolating and using unlicensed music. Splitter and it's team is not liable for any damages or misuse of Spleeter. You are responsible for your actions and we will work with music authors, artists, labels and publishers to prevent misuse and infringement of their copyrighted material if they make requests.

We assume and expect that you upload your own music only.

 

https://acondigital.com/products/acoustica-audio-editor/

Acoustica Premium Edition 7.2 — now with Stem Separation Tools

Do you know if the use the same AI for it?

Sounds similar to me

 

I want to give my insights on this subject since I've been using AI demixing for a couple years now.

Isolating instruments from mixed music using AI has been possible since mid-90s with Prosoniq Isolate. That's more than 20 years ago. Prosoniq literally pioneered the use of neural networks for audio. Then there was Audionamix ADX in 2014 and PhonicMind in 2017. You can see the pattern. It's starting to gain momentum.

One thing that makes Spleeter special is how you can train your own models by feeding your own stems to the database. You can have total control over the quality of your end result. For example: by grabbing some NI Stems from a genre like Techno, you can create your own database. Then you will be able to extract other Techno songs with more accuracy. It's not just a tool for making a capella. Plus it's open source so perhaps many experts will dive in and develop it further.
If you have some basic cmd/scripting skills, I highly suggest installing Python and grab a copy of Spleeter from the project's git repo.

splitter.ai along with moises.ai and ezstems.com are all just a frontend user interface to this technology.
I'm pretty sure none of these services are going to contribute to the models or start their own version. Perhaps they are after a subscription business model and marketing it toward novice EDM/hip hop beatmakers.


EDIT: Fixed some typos.

 

Last edited: Mar 14, 2020

They answered your question in the first paragraph. Did you read that webpage yourself?

"Acoustica 7 is the perfect solution for audio editing, mastering and restoration work on both Mac and PC. The application is available in a Premium Edition and a lower cost Standard Edition. Both editions offer a new Remix tool based on the Spleeter artificial intelligence models which can split a complete mix in up to five stems and remix them in real-time."

 

Earlier related info...
https://audiosex.pro/threads/separating-voice-piano-drums-etc-from-any-music-track-using-ai.50724/

 

@Amirious , I have little experience with the tech, beyond playing with phonicmind back in the day (didn't get much useable results out of that one), so I wanted to get your opinion on the following - what if somebody wanted to "split" not music tracks but ambience and birdcalls for video game use purposes. I got plenty of of good ambience recordings but splitting them into "brook", "seagull 1" , "seagull 2" is a time consuming pain to do. You reckon Spleeter could be tuned to do it, or it's outside the scope?

 

@Satai That's actually a great potential use case I haven't thought about before. I doubt people at Deezer had thought about it either.
Tho you would need at least more than a dozen samples of birds singing. And the longer they sing the better.

If you're in for tried and tested methods, I suggest to take a look at Spectralayers Pro. There is a steep learning curve and a couple hours of manual work but it'll be worth it.

 

Heck you can even use iZotope RX or ERA-D to split singing birds and the ambiance. Both are great pieces of software.
Just use your favorite EQ and expander to emphasize the frequencies and expand the dynamic range of singing birds vs the wind, next load one of aforementioned plugins and select a portion of the audio where birds are hyper active and let the plugin learn it as a noise profile. Now you can export the entire audio in 2 pass. Once for the ambiance then selecting output noise only to export singing birds.

 

Last edited: Mar 14, 2020

Thanks, that was the method I went for at first too, but I found that splitting based on FFT noise reduction in RX is unfortunately plagued by very annoying "underwater warbles". You can tweak settings to lessen it, but can't really make it go away entirely, so the extraction sounds unnatural as a result...

 

this splitter.ai seems to be better. Its max frequency is 16khz which is quite better than the 11khz limit of melody.ml

 

Wow... my mind is blown. I've been almost exclusively a prog rock guy the last 10 years, pretty much everything I've done has been DI/miked directly into Logic. The last year or so my writing has taken a hard turn toward ambient and electronic. I'm still wrapping my head around what Ableton can do and this splitter is already blowing my mind with possibilities. Just for my field recordings alone.

 

ezstems has that option too. It's not what you think it does tho. 16KHz doesn't necessarily increase the quality. As a matter of fact, it decreases the isolation accuracy since most of the energy of a human's voice is below the 11KHz range but as you go above, there's a shitload of frequency overlapping from synths, electric guitar and metallic bells to cymbals and fx. Also if you run the script locally, 16KHz needs more processing power than 11KHz.

 

Last edited: Mar 14, 2020

Yeah. that's what I love about the marriage of music and technology. When we stop this madness of modeling and emulating analog hardware, we can truly take advantage of all the amazing things that can only happen in digital domain.

 

ok but why?

separate recordings isn't much better?

 

Yeah but 99% of the time you won't have stems or studio quality acapella. Also, if the song you want to remix is not that popular, there are very less chances of even finding a DIY acapella.

 

Because he doesn't record the song himself.

 

Funny thing Nope
I just saw the App some days ago on JRR
The key features sounded similar
And even the names of both splitter and spleeter (same begin and end )

So my question now is evolution or ripoff ?

 

Definitely not a rip-off as they have done everything according to the license and described how they’re using Spleeter.

 

This:

and karaoke parties, practice material for singers, analyzing a mix, transcribing a composition and more future use cases we haven’t thought about yet like separation of a dialog and underlying score, resurrection of an old recording or even forensics.

 

I have tried several of these: I use Ezstems because it is the most practical: obviously the result is not exactly like having the original stems, the quality depends on the complexity of the source material.
Apparently all these stems generators have the same algorithm and therefore the same quality in recognizing and separating tracks, there doesn't seem to be one that works better than others.

a mashup created with Ezstems output: