Extracting vocals/drums/instruments from samples

[popek]

Recently, I came across an interesting technique to seperate drums/instruments etc. from samples. Maybe someone will find it useful.

Here's a link: https://github.com/deezer/spleeter/wiki
Note: it's a python library, not a vst, so it takes some time to get it working. If anyone would be interested, I can help you with setting it up.
I've tested it on the samples from the last battle and it actually works pretty well.

[HeavyDrama]

Glad you mentioned this popek. Are you a programmer?

Life is hard bro let's make it easy...
Portable version of Python and spleeter already installed with a gui https://makenweb.com/#spleetergui
Code etc here

Some FREE online alternatives
https://splitter.ai - Deezer's own online implementation. 2/5 stems split
https://www.lalal.ai/ - Vocals + Instrumental split
https://ezstems.com/ - 2/4/5 stems. Mp3 or WAV. Variable Cutoff.
https://audioalter.com/ - Vocal remover. Some other gimmicky tools that are kinda interesting.

unmix.us used to be the spot but seems dead but it's all still open-unmix

Paid alternatives that offer varying degrees of mixed audio to separate stems
Izotope RX7
Melodyne
Hit n Mix infinity
UnMixIt
Zynaptiq plugins
accusonus regroover
spectralayers
audionamix
phonicmind
virtualdj - Latest update brings real-time stems separation to any track while you're dj'ing...crazy stuff.

Pretty sure there's a bunch I'm missing will dig in my bookmarks and update this later.
I've used everything there apart from the new virtual dj and I only took the audionamix trial but if anyone needs examples before they buy let me know.

[popek]

I'm not a programmer, I'm more focused on machine learning and this kind of stuff, but I can code a bit.

The first link you posted looks promising (https://makenweb.com/#spleetergui).
I've already seen ezstems.com, but I didn't really like it because of the size limitation (up to 10MB).

The cool thing about spleeter library is that it allows you to process whole collection of samples in one click, which might be useful.

I've also tried out RX7 (free version), but in my opinion spleeter does the job better. Sounds produced by RX7 contained more artifacts, at least on the samples I used.

Btw I remember you from back in the day why don't you participate anymore?

[HeavyDrama]

I'm not a programmer, I'm more focused on machine learning and this kind of stuff, but I can code a bit.

Interesting. Would you be able to explain roughly how this works? What kind and how big a data set and are they getting that data from analysing multi-tracks etc?
I'm just wondering how difficult it would be to train a model on specific genres to get a greater accuracy and why no one has ported this into a proper vst.
I've been getting mind blowing results but my knowledge of the actual inner workings is surface level. Any help appreciated.

btw RX8 dropped recently with apparently new and improved music rebalance. I haven't had enough time to make any comparisons yet.

Batch processing is def a plus in certain situations but I've yet to need that. I find myself tweaking the settings on rx7 to get better/different results on a per track basis. Spleeter is superior for one reason, it's free. However, in terms of artifacts spectralayers has given me the best results (also using the other insane tools in there) but not by a margin that justifies it's price if that's the only feature you're interested in.

Btw I remember you from back in the day why don't you participate anymore?

A lot of the folks I used enjoy "battling" with moved on and so I just became more of a listener. Also life moves on, phases of life, new growth etc. I still make music.

peace man good to connect