Really happy that this feature arrived in rekordbox, and that it is available to me, with my DDJ400 license.
I'm not gonna say much about the quality. It feels so obvious that you probably already know yourself that this needs to be improved. And I'm sure you will do so.
The implementation is more bothering than the quality of the algorithm. And it is the main reason for me why I won't be able to use this.
Analyzing the stems only works on the fly, while a tune is loaded into one of the decks. And the results aren't saved anywhere, so I'd have to analyze the tune again, if I load it in again.
The analyzing also takes too long, and it even takes longer if I load more tunes into other decks. Because they have to be analyzed aswell, ofcourse.
I understand that analyzing takes some time, and will probably be much faster, if I had a better system. But I don't feel like that is a requirement you should set, when most people run rekordbox on a laptop.
Without the ability to analyze tunes outside of a mix and being able to save the results, I don't think any sane DJ would use this, even if you improve the quality.
I don't know anything about the technical side about how stems algorithms work, but I feel like, if you would make it 4 separations instead of 3 (like Serato does it), it might result in better quality of the individual stems.
Cette publication n’accepte pas de commentaire.
The quality is slightly lower in order to prioritize performance, even on older computers. If yours is taking a long time, what are the specs?
The reason it re-analyses every time is to prevent "size creep" of your database.
Windows 10
AMD FX-4300 Quad-Core @ 3.8 GHz
16 GB DDR3 RAM @ 1333 MHz
Nvidia GTX 1650
rekordbox and music library both being on separate SSDs
rekordbox is running totally fine on this system (without the track separation activated) Even while live-streaming to twitch via OBS from the same computer.
With track separation activated and when two tunes are on my decks, that are still being analyzed for stems, I can not even browse my library properly. And that is just with rekordbox running. I didn't dare to try it while streaming with OBS.
I understand the point of size creep. But maybe pre-hand analyzing could be done in a way that the user can analyze a selection of tracks and then have an overview somewhere that tells the user, how much space is being taken by the stems. And then just give the user the ability to clear those stems from their library at will.
Prioritize performance? Maybe I'm not understanding the point, why should quality be improvised if one has met the "basic" system requirements for proper functionality as stipulated by vendor Rekordbox?
Track selection is new but the audio quality when engaged is reminiscent of low quality early 2000 mp3 rips. I hope they improve the sound quality in future releases. Seems like not much QC went into this new function either.
RB sound quality has been consistently a cut above other competitors and one of the reasons why i have stood by the product from the get go.
Im certain that Pioneer is going to continue to work on the quality of the stems, or at least very hopeful of it. I understand the performance requirement as far too many DJs are still rocking 2014 Macbooks... but it really seems like a goofy excuse as I have not been able to find any tracks that track separation works on that would be suitable for playing on anything larger than some bedroom studio monitors. I wouldn't be caught dead playing with RB Stems at any real event because it sounds like my entire library was ripped from YouTube.
While RB continues to improve the sound quality, I bet it would be "relatively easy" for them to include a function where we could manually create stems for our tracks- ie. using more established 3rd party tools to create 3 separate layers and having Rekordbox combine them into one track with toggleable layers. Bet that'd end up being better for our low-spec friends as well.
The argument of CPU usage in Rekordbox 6.7 stems, can be solved is arbitrated as did the latest version of VirtualDJ 2023. Just add an option " low stems quality - spleeter model" for old computer, and "advanced stems quality - tensorflow enhanced model like Serato 3.0.4 stems or Virtualdj 2.0 stems". In Virtual DJ 2023 you can choose the model: old Stem 1.0 spleeter model or enhanced new Stems 2.0 own Vitual Dj model (no more spleeter...)
another solution: give the possibility to create a .mp4 audio container like traktor, with 3 audio parts in the container encoded in AAC or ALAC (optional), Vocals, drums, other, full master, in .mp4 container. This .mp4 could be produced according to specifications by an external STEMS creation software or with a batch assembler script. When Rekordbox plays a file with a single track, it stems on the fly with the fast spleeter model you already use, but if it detects an .mp4 X4 audio layers (drums, vocal, others, full master) encodes in . aac or in .alac, it no longer uses spleeter and affects the 3 tracks of the stems files in Rekordbox's multipaging audio in-memory engine. To help Rekordbox to detect .mp4 container, just add an extension: .rkb.stems.mp4, in specifications to create stems .mp4 offline container.
Like this, the user could choose to either use their existing files with the low quality stems on the fly, or play the 4 layers .mp4 container stems created before offline.
There are plenty of solutions to improve the situation, without removing what has already been done for fast, low quality, low CPU stems on the fly. Pioneer, please, listen to your loyal users, and take a look at Serato and VirtualDJ, keeping your good initial ideas.
One of the reasons for not exporting stems is for music copyright; we don't want to potentially be the big target for a music industry lawsuit.
Pulse, thanks for you answer. So make a exported saved .Stems file like Serato that only Rekordbox can read.... (protected) or keep you spleeter fast 2021 model and add an option for another model for powerfull computer. Or make the same than Serato or VirtualDJ tensorflow quality models. (Or take a look at newer model than spleeter, as fast as spleeter like GSEP)
Anyway thanks.
Yeah, the algorithm needs some work but this is like the first version. Give it time, they'll get better.
My issue is that I have a VM running a server on top of all other things. But one thing was that I maxed out my ram at 32 gigs. Plenty resources left. One thing you should do is up your ram to 32 gigs then on the extensions setting check to use more ram and uncheck multi thread. Multi thread maxes out your CPU and if you have low ram, you're going to have paging issues. In my case multi thread operation cause the track that was playing to sound garbled because the CPU was maxing out thereby cutting resources to Rekordbox. I disabled multi thread option and used more memory and the sound artifacts disappeared. Yes the track separation analysis takes a bit longer but not that much longer. I'm a happy camper :)
What Rekordbox should try is to offload the track separation analysis processing to the GPU instead of the CPU. Or in future controllers have a hardware processor for this function. Either way I know these issues will be solved in time.