Gsep (now GAudio) (2, 4, 5, 6 stem, karaoke)
Paid (20 minutes free in mp3 - no credit card required)
7$/60 minutes
16$/240 minutes
50$/1200 minutes
Electric guitar (occasionally bad), good piano, output: mp3 320kbps (20kHz cutoff), wav only for paid users, accepted input: wav 16-32, flac 16, mp3, m4a, mp4, don’t upload files over 100MB (and also 11 minutes may fail on some devices with Chrome "aw snap" error), capable of isolating crowd in some cases, and sound effects. Ideally, upload 44kHz files with min. 320kbps bitrate to have always maximum mp3 320kbps output for free.
2025 metrics for 2 stems
https://mvsep.com/quality_checker/entry/9095
“Credits are valid for 1 year from the recharge date and are used in order of earliest expiration.
Expired credits cannot be restored.”
(outdated) About its SDR
10.02 SDR for vocal model (vs Byte Dance 8.079) on seemingly MDX21 chart, but non-SDR rated newer model(s) were available from 09.06.22, and later by the end of July, and now new model is released since 6 September (there were 4 or 5 different vocal/instrumental models in total so far, the last introduced somewhere in September and no models update was performed with later UI update). MVSEP SDR comparison chart on their dataset, shows it's currently around SDR 9 for both instrumental and vocals, but I think evaluation done on demixing challenge (first model) was more precise. Be aware that GSEP causes the issue of cancelling different sounds which cannot be found in any stem.
Since the May 2024 update there was an average of 0.13 SDR increase for mp3 output and first 19 songs from multisong dataset evaluation, but judging by no audible difference for most people, they could simply change some parameters of inference. Actually, it’s more muddy now, but in some songs there are a bit less of vocal residues, and in other songs, noticeably more. Inverting the mixture with vocals in WAV will muffle the sound in overall, e.g. snares, esp. in places of these residues, but the residues will disappear as well.
Uncheck vocals to download WAV file if WAV download doesn't work,
and uncheck instrumental to download vocals in WAV -
don't check all stems if you can't download WAV at all and the download window simply disappears.
If you still can’t download your WAV files, go to Chrome DevTools>Network before starting downloading, and press CTR+R, now start download. Now both stems should be shown in DevTools>Network, starting with input file name, e.g. instrumental with ending name “result_accom.wav” (usually marked as “fail” in State column and xhr as type), click the entry with right mouse button and choose Open in new tab.
The download may fail frequently, forcing you to resume the download multiple times in browser manually, or wait a bit on the attempt to download the file at the start.
Free option of separating has been removed since the May 2024 update. There's only a 20-minute free trial with mp3 output.
Vocals and all other stems (including instrumentals/others) are paid, and length for each stem is taken from your account separately for each model.
No credit is not required for the trial.
For free, only mp3 output and 10 minutes input limit.
For paid users there's a 20 minutes limit, and mp3/wav output, plus paid users have faster queue, shareable links, and long term results storage.
Seems like there weren't any changes in the model
https://www.youtube.com/watch?v=OGWaoBOkiMg
The old files from previous separations on your account didn't get deleted so far if you have premium.
https://studio.gaudiolab.io/pricing
There was also added a new option for vocals called “Vocal remover” - good "for conservative vocals, it's fine it even has 15 best scoring on SDR." and 10.85 in vocals on multisong dataset.
Instruction
Log in, and re-enter into the link above if you feel lost on the landing page.
For instrumental with vocals, simply uncheck drums, choose vocal, and two stems will be available for download.
As for using 4/5 stem option for instrumental after mixing if you save the tracks mixed in 24 bit in DAW like Audacity, it currently produces less voice leftovers, but the instrumental have worse quality and spectrum probably due to noise cancellation (which is a possible cause of missing sounds in other stem). Use 5 stem, but cut silence in places when there is no guitar in the stem to get comparable quality to 4 stem in such places.
For 3-6 stem, you better don’t use dedicated stems mixing option - yes, it respects muting stems to get instrumental as well, but the output is always mp3 128kbps while you can perform mixdown from mp3s to even lossless 64 bit in free DAWs like Audacity or Cakewalk.
In some very specific cases you can get a bit better results for some songs by converting your input FLAC/WAV 16 to WAV 32 in e.g. Foobar2000.
Troubleshooting
- (fixed for me) Sometimes very long "Waiting" or recently “Waiting” - can disappear after refreshing the site after some time (July 2023) - e.g. if you see “SSG complete” message, you can refresh the site to change from waiting to waveform view immediately. I had that on a fresh account once when uploading the very first file on that account, and then it stopped happening (later it happened for me on an old account as well).
- (might be fixed too) If you don’t see all stems after separation (e.g. while choosing 2 stems, only vocals or only instrumental is shown) and only one stem can be downloaded (can’t be done on mobile browser) - workaround:
- "Aw snap" error on mobile Chrome can happen on regular FLACs as well as an attempt to download a song. Simply go back to the main page and try to load the song again and download it.
- If nothing happens when you press download button on PC, also go to Chrome DevTools>Network>All and click download again. Then new files will appear on the list. Right click and open mp3 file in a new tab to begin download. Alternatively, log into your account in incognito mode.
- If you have "An error has occurred. Please reload the page and try again." try deleting Chrome on mobile (cleaning cache wasn't enough in one case).
- (rather fixed) If you have “no audio” error all the time when separation is done, or preview loading is infinite, or you have only one stem, also -
In PC Chrome go to DevTools>Network>All and refresh this audio preview site, and new entries will show up on the right, which among others will list filenames with your input file name with stems names e.g. "rest of targets" in the end.
Double click it or click RBM on it and press open on new tab, and download will start.
If no filenames to download appear on the list, press CTRL+R to refresh the site, and now they should appear.
In specific cases, files in the list won’t show up, and you will be forced to log in to GSEP using incognito mode (the same account and result can be used). Also, make sure you have enough of disk space on C:.
Alternatively, clean site/browser cache (but the latter didn't help me at some point in the past, don't know how now).
If still the same, use VPN and/or new account (all three at the same time only in very specific cases when everything fails). You can also use different browser.
- When you see loop of redirections when you just logged, and you see Sign In (?~go to main page) simply enter the main link https://studio.gaudiolab.io/gsep
- If you’re getting mp3 with bitrate lower than 320kbps which is base maximum quality in this service (but you get 112/128/224 output mp3 instead)
> Probably your input file is lossy 48kHz or/and in lower bitrate than 320kbps > your file must be at least mp3 320kbps 44kHz (and not 48kHz). The same issue exists for URL option and for Opus file downloaded from YouTube when you rename it to m4a to process it in GSEP. To sum up - GSEP will always match bitrate of the input file to the output file if it’s lower than 320kbps. To avoid this, use lossless 44kHz file or if you can’t, convert your lossy file to WAV 32 bit (resample Opus to 44kHz as well - it’s always 48kHz, for YT files, don’t download AAC/m4a files - they have cutoff at 16kHz while Opus at 20kHz). Now you should get 320kbps mp3 as usual without any worse cutoff than 20kHz for mp3 320kbps.
If you still not get 320kbps, try using incognito mode/VPN/new account (at best all three at the same time).
You can use Foobar2000 for resampling e.g. Opus file (RBM on file in playlist>convert>processing>resampler>44100. And in output file format>WAV>32 bit). Don’t download from YT in any other audio than Opus, otherwise it will have 16kHz cutoff and separation result will be worse.
- (fixed) Also on mobile, the file may not appear on your list after upload, and you need to refresh the site.
- If FLAC persists to be stuck in the "Uploading" screen, try converting it to WAV (32-bit float at best)
- Check this video for fixing issues in missing sounds in stems (known issue with GSEP)
- GSEP separation results don't begin at the same time signature like UVR results.
> In order to fix it, convert mp3 to WAV or align stems manually if you need it for some comparisons or manual ensemble. Also some DAWs can correct it automatically on import.
Eventually hit their Discord server and report any issues (but they’re pretty much inactive lately).
Remarks about quality of separation
“The main difference (vs old model) is the vocals. I can't say for sure if they're better than before, but there is a difference, the "others" and "bass" are also different. Only the drums remain the same. Generally better, but the difference is not massive, depends on the song” (becruily)
GSEP is generally good for tracks where using all the previous methods you had bleeding (e.g. low-pitched hip-hop vocals) or got flute sounds removed, although it struggles with “cuts” and heavily processed vocals in e.g. choruses. Though, it has more bleeding in some cases when the very first model didn't, so new MDX-UVR models can achieve generally better results now.
"GSEP is good at piano extraction, but it still lacks in vocal separation, in many times the instruments come out together with the voices, this is annoying sometimes."
Electric guitar model got worse in the last update in some cases. Also, bass & drums also not so loud since the first release of gsep.
"Electric guitar model barely picks up guitars, it doesn't compare to Demix/lalal/Audioshake".
“I kinda like it. When it works (that's maybe 50-60% of time), it's got merit.”
The issue happens (also?) when you process (GSEP) instrumental via 5 stems. If you process a regular song with vocals - it picks up guitar correctly. It happens only in a place where previously was vocal removed by GSEP 2 stem.
I only tested GSEP instrumental so far, I don’t know whether it happens on official instrumentals too (maybe not).
The cool thing is that when the guitar model works (and it grabs the electric), the remaining 'other' stem often is a great way to hear acoustic guitar layers that are otherwise hidden.
The biggest thing I'd like to see work done on is the bass training. At present, it can't detect the higher notes played up high... whereas Demucs3/B can do it extremely well.”
It has “much superior” other stem than Demucs or even better than Audioshake. It has changed since 6 September 2022, but probably got updated since then and is probably fine.
As for 14.10.22 piano model sounds “very impressive”.
As for the first version of the model comparable vocal stem to MDX-UVR 9.7, but with current limitation to mp3 320kbps and worse drums and bass than Demucs (not in all cases). Usually less bleeding in instrumentals than VR architecture models.
“Gsep sounds like a mix between Demucs 3 and Spleeter/lalal, because the drums are kind of muffled, but it's so confident when removing vocals, there aren't as many noticeable dips like other filtered instrumentals, and it picks up drums more robustly than Demucs. [it can be better in isolating hihats then Demucs 4 ft model too]
It removes vocals more steadily and takes away some song's atmospheres, rather than UVR approach which tries to preserve the atmosphere, but [in UVR] you end up with vocal artefacts”
As for tracks with more complicated drums sections: “GSEP sounds much fuller, Demucs 3 still has this "issue" with not preserving complex drums' dynamics” it refers to e.g. not cancelling some hi-hats even in instrumentals.
It happens that some instruments can be deleted from all stems. “From what I've heard, [it] gets the results by separating each stem individually (rather than subtractive / inverting etc.), but this means some sounds get lost in between the cracks you can get those bits by inverting the gsep stems and lining up with the original source, you should then be left with all the stuff gsep didn't catch”.
Also, I'd experiment with the result achieved with Demucs ft model, and apply inversion for just the specific stem you have your sounds missing.
As for June 2023 gsep is still the best in most cases for stems, not anywhere close to being dead
gsep loves to show off with loud synths and orchestra elements, every other mdx/demucs model fail with those types of things
Processing
After your track is uploaded (when 5 moving bars disappear) it’s very fast, and it takes 3-4 minutes for one track to be separated using 2 stem option (processing takes around 20 seconds). If 5 bars are moving longer than expected track upload time, and you see that nothing uses your internet upload, simply press CTRL+R and retry, if still the same, log off and log in again. It can rarely happen that the upload stuck (e.g. when you minimize the browser on mobile or switch tabs).
Generally it’s very fast and long after the very first GSEP days, I needed to wait briefly in queue twice at 6-9 PM CEST, and I think once on Sunday in weekend of adding new model once in my whole life I waited around 7 minutes. Usually you wait in a queue longer than processing takes, so it’s bloody fast.
___
(outdated)
If your stems can’t be downloaded after you click the download button, go to Tools for Developers in your browser and open the console and retry. Now you should see an error with file address and your file name in it. You can simply copy the address to the address bar and start downloading it.
(Outdated - 3rd model changes) The quality of hi-hats is enhanced, sometimes at the cost of less vivid snare in less busy mix, while it’s usually better in busy mix now, but it sometimes confuses snare in tracks when it sounds similar to hi hat making it worse than it was. So a trap with lots of repetitive hi-hats and also tracks with a busy mix should sound better now.