I use stem separation for mixing live projects. It's absolutely awesome at clearing up spill in vocal mics. A mixing project I've just finished (currently exporting right now), I used stem separation to split the vocal elements of the drum overheads so I could automate them up for audience segments, as they didn't have dedicated audience mics at the live event. It's one of the craziest technologies to arrive, but SO many great uses. I use SpectraLayers Pro 10 here. EDIT: P.S. Enable Graphics Card decoding if you use an Nvidia card. It's so much faster, even without a top-end card. Much faster than CPU at splitting stems and other options.
Now that took some putting together but undoubtedly this is the definitive guide to stem extraction. The results were very interesting and like you I was amazed how clear the AI-Lid stems were. I use spectralayers in cube to split vocals or for use on my reproduction tracks. They are usually "good enough" because they will be played back over my versions of the originals but certainly, as you have proved, would not stand up to being solo'ed.
Fantastic demo on this, very helpful. I did a lot of music production up until about 10 years ago, and didnt know this was possible, until about 4 months ago.... I am absolutely blown away that technology has come this far, and kind of made a fool of myself finding out. There's a friend on Facebook who kept posting audio samples that he had ripped... I thought he was just playing acapellas but, one day noticed that he had all of the separate tracks... I ask what it was about, and he told me that he had split all the tracks... 'what from an original record?" Yes, was his reply... To which I replied, "Thats impossible", as I thought he just had the multitrack of the record... He then proved it to me by separating a track of my choice... I could not apologise more!!! Never in my life would i ever of thought this was possible... I am not sure if the software he uses is one you have shown, as there is only a few people with his and it is Unix based. What I can tell you is, this split track was virtually perfect... It was that incredible, I almost filled up hahaha... The best one you used on your demo IMO was AI Lyd.. It was amazing, this person on Face Book, I highly recommend as I think it is almost indistinguishable from an original track. As you pointed out, it depends on the parts etc but, everything I here by him is AI Lyd and then some. I used to produce a lot of sampled based dance music years ago... OMG, if this technology, or similar, would of been around then... I'd of been very rich and in absolute heaven!!! Thanks for a great video.
Thanks for sharing, Les! I don't think you'll be alone in this at all. I regularly speak to people who don't know that this is possible; not everyone spends their time following the ins and outs of audio software development, and there are plenty of people who record music who are in their 'bubble', and know what they know (and plenty of it) but are unaware of what is happening with these sorts of developments. It'd be interesting to see the specific software that's being used if you have any way of finding out.
And for my 4th comment after watching: a nice tip for all of you. Soulseek is alive and strong, and it is the best way to get good quality audio files (even HI RES WAV - FLAC at 24 bits and even 48 & 96 kHz. Give it a go!!! Feeding good quality audio into these splitters really makes a difference!!
It's still pretty amazing that you are able to actually split each instrument apart from a complete mix, This is pretty ok for taking out some riffs to be used in a backing track. The sound quality is very ok for sampling some tracks that you can't do live as 1 keyboard guy when you don't have a full brass and string section to or some very specific sounds that you just can't be bothered with trying to replicate on your instrument of choice
I’ve only ever used Izotope RX Advanced for doing stems. It only gives you 4 stems but on it’s “best” setting I’ve usually got enough of what I need. Great research as usual Darren,
Just tried Demucs GUI on a project I've been trying to remix for a few years now. The results are noticeably superior to SpectraLayer, RX, and the others I've tried. The separation is more complete. Demucs GUI is a keeper. Thanks so much! 🍻
I would not try to use any of the present online or offline stem tools to get near perfect stems. It doesn't even work well enough for mixes with few tracks and well audible separation - and it is more problematic for all I tested with complex mixes involving various keyboards/synths. BUT: Rip-X (for transcription of songs and for analyzing and manipulating audio) and Spectralayers (for analyzing and manipulating audio) are fantastic analytic and editing tools while separating - both significantly ahead of any of the others named here IMO. On the other hand, you are right that Demucs seems to do the best job in stem separation without too much overlap and phasing, and it is suprisingly fast while doing such a good job.
I knew I'd missed something; I had quite a lot of time pressure to get the video out and it took days to make it so apologies for that. RipX's bass was indeed much the same, but again, sorry for the error!
I don't know how you did this, I have never watched a 90min YT vid before. SO qdos to you, I came here after spotting the Demucs vid you did and you kept my normally short attention span rolling. Would be interested in the audio to midi vid. And also for us hobby only strummers, any info on 'free' voice / instrument replacement / song creation etc etc AI tools would be great. Keep up the good work, Kind Rgds.
Thanks Joe - much appreciated. I did wonder how many people would watch the entire thing, but thought it better that everything was in there and people can pick and choose... And thanks for the suggestions, they're on the list.
Yes, demucs v1,2,3 and 4 are options. I wasn't aware that it did full stem output (I thought it only did minus vocals), so every day is a school day! Thanks.
@@louminati4318Yeah, I'm planning on making a separate video on audio to MIDI conversion services and software as it's a whole other topic on its own and there are a lot of options (as here!).
I managed to separate drums (kick, snare, toms & cymbals from a stereo drum recording, and I was amazed with the results. It's a specially trained model derived from Demucs Hybrid that can be loaded on both UVR and Demucs GUI. The sky is the limit with this AI stuff!!
Nice video... But I have a very hard time defining that track as 'rock' 😄 I would love to see same comparison with (extreme) metal... I bet you would get very very different results, even from the best 🙂
Brilliant. I was wondering if there is an indication of cost with Ai-Lyd? I have used Stemroller but that is clearly a ways behind some of the ones here.
I'm going to do a separate video on this as this one was already long enough and I thought it was different enough to warrant that. It'll be a few weeks though as these take a long time to make!
Hi Ive got a 24 track home recording studio and listened to to the best stem splitters. .the free one Demucs Gui and AI Lyd worked out the best and I agreed with you.. I cant seem to find the AI Lyd to down load any ideas.. The free one even your easy way to download it was a bit complicated for this old guy.. So I need help with the AI Lyd cheers Frankie.. Excellent presentation by the way best I've seen
@@musictechtuition Went on there but totally confuses me. All I want to do is use their stem stripping software as you did in your presentation for my songs but cant see how to do it..With other sites you just load in a song and use their stem remover to test it why cant it be done with this company?
@@frankieAllan-vr6qn As I understand it, they fine-tune the splitting process for each track, so you need to fill in the form on the site and upload the file and then they will take it from there. They do a free sample service (a small part of the track you send) so you can see if it'll work for you...
One of my favorite uses of these software is to isolate vocals from a bands live recording. No more battling drum bleed with gates... I just run my lead and backing vocal microphones through these things and I have near-studio clean vocal tracks that I can use compression on without getting the good old "SCHHCH" from cymbals
Thanks for this, great info and could save a few bucks, as I've been hoping for Spectralayers 10 to be discounted before plunging in.. One question these reviews always seem to skirt/avoid/ignore, though, is how well these tools work at separating other sounds. Most of my old tracks feature synths, pianos, sax, etc, often more than six elements at a time. Can these systems do anything with that, or does it all get categorised as 'other'?
At the moment, I think 'other' is everything that isn't extracted, which can vary with tracks/sounds. I've had some times when 'other' has included parts of the named stems (i.e. a bit of guitar being on the 'other' track). I know there will be options in the future to do multiple runs over a track (in the way I think that ai-lyd splits drums out - I think they get the drum stem and then from there they can break it down further), and I'm looking to see if it's possible to do it with something like demucs-GUI, but it's dependent on appropriate models and I've not spent enough time finding ones which will do the trick.
AI-lyd - a small company on a small island on the Norwegian west cost, halfway between Bergen and Trondheim. Lyd means sound in Norwegian and they're taking great strides in stem splitting etc. It seems the competition will have to up their games?
According to deezer's github [1], RX8 was using spleeter, so I'd expect the performance to be the same. It may no longer use it (I can't find any open source licence details on their site, so they may well no longer be using it, but I don't own a licence to a full version of RX to test it. [1] - github.com/deezer/spleeter
I am on the rip x daw pro camp but prefered the old version. And with rip x you get free lifetime upgrades and yes the artifacts but those artifacts were probably in the original recording and yes it's not perfect but you can literely demix an entire track but if you choose the layers option which I do, it leteraly picks up everything like drums that are not lined up properly to the effects on the vocals. And for an example I ripped an early van halen track and sure enough it picked up all the instrument bleeds that were actaully in the original recordings. Or I ripped the entire helloween mini ep and thier first album walls of jericho and it literely picked up that Ingos drums were not aligned corectly as he stated in interviews that on the first 2 albums they knew nothing about recording then especially with drums. Anyway you can demix an entire track with rip x down to seperating everything down to each individual thing cymbals. Tom's down to all the effects. And it has a, keyboard.
The only practical use for me would be splitting out drums for a nice usable drum pattern starting point, and using a DRUM REPLACER to get the MIDI of it and modify it all to purpose and of course just skip all the quality issues. I expect it won't be long, and we can simply directly get the MIDI and bypass the stem step completely via AI. In fact, I expect AI will just spit out alternate takes based upon that starting point, or even the ability to merge TWO drum "stems" (again, not really getting the stems as an interim step, but directly from two (or more) full mixes of songs) and have AI produce a MIDI drum track that is a random BLEND of those two patterns, sort of like MORPHING between them for semi-infinite variety with some knobs or sliders to vary the final output... We should shortly be going beyond what ReMIDI can do already in the MIDI domain and be able to use ALL recordings as sources to get not just audio quality limited stems shown here, but good quality MIDI from ANY INSTRUMENT in a full mix for repurposing. THAT will be fun.
They may have run out of money to pay their cloud processor (AWS/Google Cloud or whoever) - it's not free to do the sort of thing that's done by these services, so I've seen others that have gone this way for the same reason, and usually these things are done by people as a passion project or using a free allocation of processing which when it runs out, the system goes offline.
Some of the services only took compressed files as inputs, and it was a 320kbps MP3 and in the blind tests I've seen they are not reliably distinguishable from uncompressed audio. Certainly if you do a null against the MP3 and the original wav in many cases there's very little remaining.
It's horrible to use and buggy as heck workflow has alot of hidden main options that are unintuitive so they have a long way to go and a change of UI design change which they may never do
They just did an update to rip x with some changes to the interface I would be happy if the brought back the highest quality option or the slider for quality like it used to have. @@williamshaneblyth
Yes, but...no. It has A LOT of extra features, and I've tested it extensively, and I really think I should be getting way better results. I love the amount of control it gives you (AFTER splitting) but the model/algorithm they're using is not among the best. Not even close. JMHO. Cheers!
The complete idea of splitting each and every second of audio is bad.. why not fishing the best sounds by each instrument then baking the arrangement together. Sure its not possible with voices.. but the rest should work well with the new AI stuff.
Dmucs gui is such a win!. Thanks for the video 🔥
Dmucs isn't working for me recently. Anyone else facing this?
@@MLATX512shucks I’ve been using spectralayers sorry , they have a long trial period so worth a shot
I use stem separation for mixing live projects. It's absolutely awesome at clearing up spill in vocal mics. A mixing project I've just finished (currently exporting right now), I used stem separation to split the vocal elements of the drum overheads so I could automate them up for audience segments, as they didn't have dedicated audience mics at the live event.
It's one of the craziest technologies to arrive, but SO many great uses. I use SpectraLayers Pro 10 here.
EDIT: P.S. Enable Graphics Card decoding if you use an Nvidia card. It's so much faster, even without a top-end card. Much faster than CPU at splitting stems and other options.
Waiting for the Demucs GUI video now!
Now that took some putting together but undoubtedly this is the definitive guide to stem extraction. The results were very interesting and like you I was amazed how clear the AI-Lid stems were. I use spectralayers in cube to split vocals or for use on my reproduction tracks. They are usually "good enough" because they will be played back over my versions of the originals but certainly, as you have proved, would not stand up to being solo'ed.
Fantastic demo on this, very helpful.
I did a lot of music production up until about 10 years ago, and didnt know this was possible, until about 4 months ago.... I am absolutely blown away that technology has come this far, and kind of made a fool of myself finding out. There's a friend on Facebook who kept posting audio samples that he had ripped... I thought he was just playing acapellas but, one day noticed that he had all of the separate tracks... I ask what it was about, and he told me that he had split all the tracks... 'what from an original record?"
Yes, was his reply... To which I replied, "Thats impossible", as I thought he just had the multitrack of the record... He then proved it to me by separating a track of my choice... I could not apologise more!!! Never in my life would i ever of thought this was possible... I am not sure if the software he uses is one you have shown, as there is only a few people with his and it is Unix based. What I can tell you is, this split track was virtually perfect... It was that incredible, I almost filled up hahaha...
The best one you used on your demo IMO was AI Lyd.. It was amazing, this person on Face Book, I highly recommend as I think it is almost indistinguishable from an original track. As you pointed out, it depends on the parts etc but, everything I here by him is AI Lyd and then some.
I used to produce a lot of sampled based dance music years ago... OMG, if this technology, or similar, would of been around then... I'd of been very rich and in absolute heaven!!!
Thanks for a great video.
Thanks for sharing, Les! I don't think you'll be alone in this at all. I regularly speak to people who don't know that this is possible; not everyone spends their time following the ins and outs of audio software development, and there are plenty of people who record music who are in their 'bubble', and know what they know (and plenty of it) but are unaware of what is happening with these sorts of developments. It'd be interesting to see the specific software that's being used if you have any way of finding out.
And for my 4th comment after watching: a nice tip for all of you. Soulseek is alive and strong, and it is the best way to get good quality audio files (even HI RES WAV - FLAC at 24 bits and even 48 & 96 kHz. Give it a go!!! Feeding good quality audio into these splitters really makes a difference!!
Thanks Darren 🙂 For this great comparison video. cheers
It's still pretty amazing that you are able to actually split each instrument apart from a complete mix, This is pretty ok for taking out some riffs to be used in a backing track. The sound quality is very ok for sampling some tracks that you can't do live as 1 keyboard guy when you don't have a full brass and string section to or some very specific sounds that you just can't be bothered with trying to replicate on your instrument of choice
I’ve only ever used Izotope RX Advanced for doing stems. It only gives you 4 stems but on it’s “best” setting I’ve usually got enough of what I need. Great research as usual Darren,
Just tried Demucs GUI on a project I've been trying to remix for a few years now. The results are noticeably superior to SpectraLayer, RX, and the others I've tried. The separation is more complete. Demucs GUI is a keeper. Thanks so much! 🍻
Highly unlikely it's superior to SpectraLayers 10 as they use the same AI model. RX uses an older model though (Spleeter-based).
I would not try to use any of the present online or offline stem tools to get near perfect stems. It doesn't even work well enough for mixes with few tracks and well audible separation - and it is more problematic for all I tested with complex mixes involving various keyboards/synths. BUT: Rip-X (for transcription of songs and for analyzing and manipulating audio) and Spectralayers (for analyzing and manipulating audio) are fantastic analytic and editing tools while separating - both significantly ahead of any of the others named here IMO. On the other hand, you are right that Demucs seems to do the best job in stem separation without too much overlap and phasing, and it is suprisingly fast while doing such a good job.
Excellent mate 👏👏👏🎶🥂✨
Seems like you played Dance Bass twice for Fadr, leaving RipX DAW aside. Not that it changes much though, I suspect ;)
I knew I'd missed something; I had quite a lot of time pressure to get the video out and it took days to make it so apologies for that. RipX's bass was indeed much the same, but again, sorry for the error!
Hey great video! Did you cover the MIDI extraction features in any other video, as mentioned at the beginning of this video?
I don't know how you did this, I have never watched a 90min YT vid before. SO qdos to you, I came here after spotting the Demucs vid you did and you kept my normally short attention span rolling. Would be interested in the audio to midi vid. And also for us hobby only strummers, any info on 'free' voice / instrument replacement / song creation etc etc AI tools would be great. Keep up the good work, Kind Rgds.
Thanks Joe - much appreciated. I did wonder how many people would watch the entire thing, but thought it better that everything was in there and people can pick and choose... And thanks for the suggestions, they're on the list.
Hi isn't Demucs an option on UVR5?
Yes, demucs v1,2,3 and 4 are options. I wasn't aware that it did full stem output (I thought it only did minus vocals), so every day is a school day! Thanks.
@@musictechtuition FADR also gives midi parts for tracks. It’s not perfect, neither am I but I have a bit of fun using Cubase 12.
@@louminati4318Yeah, I'm planning on making a separate video on audio to MIDI conversion services and software as it's a whole other topic on its own and there are a lot of options (as here!).
I managed to separate drums (kick, snare, toms & cymbals from a stereo drum recording, and I was amazed with the results. It's a specially trained model derived from Demucs Hybrid that can be loaded on both UVR and Demucs GUI. The sky is the limit with this AI stuff!!
Could you do a new version of this video but comparing the Steingberg Stems separation vs Logic vs Fl Studio and Studio One??
Nice video... But I have a very hard time defining that track as 'rock' 😄
I would love to see same comparison with (extreme) metal... I bet you would get very very different results, even from the best 🙂
Brilliant. I was wondering if there is an indication of cost with Ai-Lyd? I have used Stemroller but that is clearly a ways behind some of the ones here.
The costs are on their webpage ar around 13:56
@@musictechtuitionD'oh! my bad
@@goport No worries!
Brilliant comparison, thx. Would like to know if you had a look for the MIDI and Lyrics that was checked on Spleeter. Was it useful?
I'm going to do a separate video on this as this one was already long enough and I thought it was different enough to warrant that. It'll be a few weeks though as these take a long time to make!
Hi Ive got a 24 track home recording studio and listened to to the best stem splitters. .the free one Demucs Gui and AI Lyd worked out the best and I agreed with you.. I cant seem to find the AI Lyd to down load any ideas.. The free one even your easy way to download it was a bit complicated for this old guy.. So I need help with the AI Lyd cheers Frankie.. Excellent presentation by the way best I've seen
Thanks Frankie. Ai-Lyd is online only - it's a web service - so you'll need to visit them at ai-lyd.no/en/
@@musictechtuition Went on there but totally confuses me. All I want to do is use their stem stripping software as you did in your presentation for my songs but cant see how to do it..With other sites you just load in a song and use their stem remover to test it why cant it be done with this company?
@@frankieAllan-vr6qn As I understand it, they fine-tune the splitting process for each track, so you need to fill in the form on the site and upload the file and then they will take it from there. They do a free sample service (a small part of the track you send) so you can see if it'll work for you...
@@musictechtuition Hi did this but where can I listen to the result..2 days since i uploaded a track
@@frankieAllan-vr6qn They will contact you with the result; when I sent the track to them in the video, everything was done via email from them.
One of my favorite uses of these software is to isolate vocals from a bands live recording.
No more battling drum bleed with gates... I just run my lead and backing vocal microphones through these things and I have near-studio clean vocal tracks that I can use compression on without getting the good old "SCHHCH" from cymbals
Awesome
Thanks for this, great info and could save a few bucks, as I've been hoping for Spectralayers 10 to be discounted before plunging in.. One question these reviews always seem to skirt/avoid/ignore, though, is how well these tools work at separating other sounds. Most of my old tracks feature synths, pianos, sax, etc, often more than six elements at a time. Can these systems do anything with that, or does it all get categorised as 'other'?
At the moment, I think 'other' is everything that isn't extracted, which can vary with tracks/sounds. I've had some times when 'other' has included parts of the named stems (i.e. a bit of guitar being on the 'other' track). I know there will be options in the future to do multiple runs over a track (in the way I think that ai-lyd splits drums out - I think they get the drum stem and then from there they can break it down further), and I'm looking to see if it's possible to do it with something like demucs-GUI, but it's dependent on appropriate models and I've not spent enough time finding ones which will do the trick.
@@musictechtuition Thanks for the kind and detailed reply. I will have to explore these tools, and if I find out anything more I'll come back to you.
AI-lyd - a small company on a small island on the Norwegian west cost, halfway between Bergen and Trondheim. Lyd means sound in Norwegian and they're taking great strides in stem splitting etc. It seems the competition will have to up their games?
I think so - hopefully we'll all eventually have access to such great-sounding stems in-app but it's good to know you can get them already from them.
UVR 5 with Demucs
Great review! But am I the only person who finds it deeply offensive that the splitters relegate guitars to the category, "Other"?
Nope! I've played the guitar since I was 13... so I'm equally offended!
Interesting. So what does the first song sound like when remixing the stems together?
I would also like to hear how well each of these were able to rebuild the song from the stems.
wowzers, one of the few people on youtube who know how to pronounce et cetera. not 'ex cetera' , that's right!
OK, I'll own up... I went to a school where you had to do a year of Latin. And this is all I have to show for it!
what about izotope RX MusicRebalance?
According to deezer's github [1], RX8 was using spleeter, so I'd expect the performance to be the same. It may no longer use it (I can't find any open source licence details on their site, so they may well no longer be using it, but I don't own a licence to a full version of RX to test it.
[1] - github.com/deezer/spleeter
crap
I am on the rip x daw pro camp but prefered the old version.
And with rip x you get free lifetime upgrades and yes the artifacts but those artifacts were probably in the original recording and yes it's not perfect but you can literely demix an entire track but if you choose the layers option which I do, it leteraly picks up everything like drums that are not lined up properly to the effects on the vocals. And for an example I ripped an early van halen track and sure enough it picked up all the instrument bleeds that were actaully in the original recordings.
Or I ripped the entire helloween mini ep and thier first album walls of jericho and it literely picked up that Ingos drums were not aligned corectly as he stated in interviews that on the first 2 albums they knew nothing about recording then especially with drums.
Anyway you can demix an entire track with rip x down to seperating everything down to each individual thing cymbals.
Tom's down to all the effects. And it has a, keyboard.
Which model does Moises use?
They started with spleeter, but then started adding their own stuff
The only practical use for me would be splitting out drums for a nice usable drum pattern starting point, and using a DRUM REPLACER to get the MIDI of it and modify it all to purpose and of course just skip all the quality issues. I expect it won't be long, and we can simply directly get the MIDI and bypass the stem step completely via AI. In fact, I expect AI will just spit out alternate takes based upon that starting point, or even the ability to merge TWO drum "stems" (again, not really getting the stems as an interim step, but directly from two (or more) full mixes of songs) and have AI produce a MIDI drum track that is a random BLEND of those two patterns, sort of like MORPHING between them for semi-infinite variety with some knobs or sliders to vary the final output...
We should shortly be going beyond what ReMIDI can do already in the MIDI domain and be able to use ALL recordings as sources to get not just audio quality limited stems shown here, but good quality MIDI from ANY INSTRUMENT in a full mix for repurposing. THAT will be fun.
Spleeter Online is... Offline? Anyone know why? and apologies if this has been asked previously by someone else.
They may have run out of money to pay their cloud processor (AWS/Google Cloud or whoever) - it's not free to do the sort of thing that's done by these services, so I've seen others that have gone this way for the same reason, and usually these things are done by people as a passion project or using a free allocation of processing which when it runs out, the system goes offline.
AI-Lyd sounds like the AudioShake algo...
Hi. It's NOT.🙂
Imo, if the goal is the critical listening, the input should not be an mp3. For comparison maybe, but anyway...
Some of the services only took compressed files as inputs, and it was a 320kbps MP3 and in the blind tests I've seen they are not reliably distinguishable from uncompressed audio. Certainly if you do a null against the MP3 and the original wav in many cases there's very little remaining.
If a stem splitter service only allows for mp3 uploading, for me is a big NO even before trying 😂
Thanks for the comparison, I REALLY appreciate the effort!!
Если я правильно понял, то соло гитару никто никогда не стемзил.(успешно )Соло гитара не поддается. Очень жаль.
UVR5
ai=artificial intelligence (you knew that :)
lyd=sound in norwegian
.no=norwegian server
correct ! :-)
RipX has been known for being the best one for a long time
It's horrible to use and buggy as heck workflow has alot of hidden main options that are unintuitive so they have a long way to go and a change of UI design change which they may never do
Not anymore, even moises is better.
They just did an update to rip x with some changes to the interface I would be happy if the brought back the highest quality option or the slider for quality like it used to have. @@williamshaneblyth
Nah RipX sucks. UVR , demucs is best by far
Yes, but...no. It has A LOT of extra features, and I've tested it extensively, and I really think I should be getting way better results. I love the amount of control it gives you (AFTER splitting) but the model/algorithm they're using is not among the best. Not even close. JMHO. Cheers!
You left out using Moises.
Lyd is sound in Norwegian 😂
I really should have checked that before recording.... Noted for the future!
The first two were the same.
The complete idea of splitting each and every second of audio is bad.. why not fishing the best sounds by each instrument then baking the arrangement together.
Sure its not possible with voices.. but the rest should work well with the new AI stuff.
the website doesnt work!
Which one?
@@musictechtuition Having some difficulties with webserver. Sorry! Ai-Lyd CEO