Ripple on Windows or MacOS
- Another way to use Ripple without Apple device -
virtual machine
Sideloading of this mobile iOS app is possible on at least M1 Macs.
- Saucelabs
Sign up at https://saucelabs.com/sign-up
Verify your email, upload this as the IPA: https://decrypt.day/app/id6447522624/dl/cllm55sbo01nfoj7yjfiyucaa
Rotating puzzle captcha for TikTok account can be tasking due to low framerate. Some people can do it after two tries, others will sooner run out of credits, or completely unable to do it.
Mobile device cloud
- Scaleway
"if you're desperate you can rent an M1 Mac on scaleway and run the app through that for $0.11 an hour using this https://github.com/PlayCover/PlayCover”
IPA file:
https://www.dropbox.com/s/z766tfysix5gt04/com.ripple.ios.appstore_1.9.1_und3fined.ipa?dl=0
"been working like a dream for me on an M1 Pro… I've separated 20+ songs in the last hour"
More info:
-https://cdn.discordapp.com/attachments/708579735583588366/1146136170342920302/image.png
- “keep in mind that the vm has to be up for 24 hours before you can remove it, so it'll be a couple bucks in total to use it”
Fixing chunking artefacts (probably fixed)
- Every 8 seconds there is an artifact of chunking in Ripple. Heal feature in Adobe Audition works really well for it:
https://www.youtube.com/watch?v=Qqd8Wjqtx-8
-The same explained on RX10 example and its Declick feature:
https://www.youtube.com/watch?v=pD3D7f3ungk
Volcengine (a.k.a. The sami-api-bs-4track - 10.8696 SDR Vocals)
https://www.volcengine.com/docs/6489/72011
Ripple/SAMI Bytedance's API was found. If you're Chinese, you can go through it easier -
you need to pass the Volcengine facial/document recognition, apparently only available to Chinese people
We already evaluated its SDR, and it even scored a bit better than Ripple itself.
"API from volcengine only return 1 stem result from 1 request, and it offers vocal+inst only, other stems not provided. So making a quality checker result on vocal + instrument will cost 2x of its API charging.
Something good is that volcengine API offers 100 min free for new users"
API is paid 0.2 CNY per minute.
It takes around 30 seconds for one song.
It was 1.272 USD for separating 1 stem out MVSEP's multisong dataset (100 tracks x 1 minute).
"My only thought is trying an iOS Emulator, but every single free one I've tried isn't far-fetched where you can actually download apps, or import files that is"
So far, Ripple didn't beat voc_ft (although there might be cases when it's better) and Dango.
Samples we got months ago are very similar to those from the app, also *.models files have SAMI header and MSS in model files (which use their own encryption), although processing is probably fully reliable on external servers as the app doesn't work offline (also model files are suspiciously small - few megabytes, although it's specific for mobilenet models). It's probably not the final iteration of their model, as they allegedly told someone they were afraid that their model will leak, but better than the first iteration judging by SDR with even lossy input files.
Later they told that it’s different model than the one they previously evaluated, and that time it was trained with lossy 128kbps files due to some “copyright issues”.
"One thing you will notice is that in the Strings & Other stem there is a good chunk of residue/bleed from the other stems, the drum/vocal/bass stems all have very little to no residue/bleed" doesn't exist in all songs.
It's fully server-based, so they may be afraid of heavy traffic publishing Ripple worldwide, and it's not certain whether it will happen.
Thanks to Jorashii, Chris, Cyclcrclicly, anvuew and Bas, Sahlofolina.
Press information:
https://twitter.com/AppAdsai/status/1675692821603549187/photo/1
https://techcrunch.com/2023/06/30/tiktok-parent-bytedance-launches-music-creation-audio-editing-app/
Site:
BS-RoFormer
Used architecture in Capcut/Ripple (now defunct). Their paper was published and later reimplemented by lucidrains for training and inferencing:
https://github.com/lucidrains/BS-RoFormer
Later, Mel-Band RoFormer based on band split was released, which is faster, but doesn't provide such high SDR as BS. Mel variant might require some revision of the code, and its paper might lack some features need to keep up SDR-wise with extremely slow BS original variant. On paper, it should be better than BS-Roformer, but for some reason, models trained with Mel have worse results than with BS-Roformer (so probably problem with reimplementation from paper). Kim reworked her config, so the results with Mel models improved, but still are a tad lower than BS-Roformer. ZFTurbo includes training and inference of Roformers in his repository on GitHub.
For more information, check the training section.
About ByteDance
Winners of MDX23 competition. They said at the beginning, that it utilizes novel arch (so no weighting/ensembling of existing models). In times of v.0.1 seemingly the best vocals, not so good instrumentals, as it was once said by someone who heard samples, but they came a long way lately. It's all about their politics. It's a Chinese company responsible for TikTok, famous for d**k moves outside China - manipulating their algorithms - encourage of stupidity outside China, and greedy, wellness-centered attitudes for users in China (the app is currently banned in China), manipulating their algorithms to promote only black-white relationships in western countries, spying on users copying their clipboard, spying even on journalists to find their sources of information about the company, unauthorized remote access to TikTok user data from China, and also, a subject to ban in US and other countries for bad influence on children, data infringement by storing non-China users data directly on their servers which is against the law of many countries (there were some actions taken on it later). Decompiling TikTok analysis (tons of spying improper behavior of the app). Currently, Bytedance is only around 40% owned by founders, Chinese investors, and their employees and the rest (60%) state global investors (incl. lots of American) and is pushed to sale more stakes to US risking US ban on the app.
They said, the CEO, told them to hold this ByteDance arch for two years for themselves. Initially they had plans to release it in some kind of app, firstly at the end of June, later something was planned at the end of year, later they said something about two years (maybe more about open sourcing, but we can't have our hopes high). Previously, they said the case of open sourcing/releasing was stuck in their legal department. Later they told they used MUSDBHQ+500 songs for their dataset. These 500 songs could have been obtained illegally for training (although everyone does it), but they might be extremely cautious about it (or it's just an excuse). Eventually, they released Ripple and Capcut. Then they released the paper for the Bs/Mel archs, and it was implemented and coded by Lucidrains, so later could be used for training.
Later, they seemingly spread information among users privately, that despite the similarities in SDR, the 18.75 score is a result of a trolling, someone other than ByteDance. Some people favoring ByteDance were rumored for disruptive, trolling behavior on our server too, harassing other users, or just being unkind to others etc. Besides, the same person responsible, was also the most informed about ByteDance next moves, and was also changing nicknames or accounts frequently. Also possessed great ML knowledge. Many coincidences. In the end, the same user, zmis (if you see the details of the account above), was behind a lot of newly created, accounts, which were banned on our server.
The same day or in very similar period, a new account was created, conducting the same behavior, when previous was banned.
The main core of their activities, was spreading misinformation about SDR metrics, telling that is the most important thing in the world, because their own arch is good at it, hence the narration.
So don't bother, and do your good job not feeding troll from other company. They don't like competition, doing their own moves behind backdrops and become better.
It’s not impossible to fake SDR results in the MVSEP leaderboard. For current public archs, you’d need to feed your dataset both by the songs in the evaluation dataset, keeping your regular big dataset in place, so you simply lose evaluation factor of this leaderboard, or you can simply mix your result stems with original stems. “SDR focuses more on lower frequencies, it can easily be fooled into giving a higher score if the lower frequencies are louder, Bas tested this theory and confirmed it”. you can boost the bass, and it will score +1 or +2 sdr higher or something, that's why It's not always reliable” - becruily
Those results, which are not faked, are at least those, which were uploaded by various users evaluating the same public, available for offline use models, but usually uploaded with various parameters which affects SDR (so usually the better parameters, the higher SDR, but not always), remain consistent among various users evaluations with similar parameters and inference code, so scalability is correct and preserved, thus the results weren’t faked, and can be reproduced with similar SDR. For the other scores from unpublic inferences/models/methods, we simply trust ZFTurbo and rather viperx too, as they’re/were our trusted users for years. Also, the leaderboard in the current multisong dataset tends to give better SDR to the results with more residues on different occasions before, so the chart is simply not fully reliable for that, but rather not manipulated in its core either. It’s more a nature of SDR measurement and/or used dataset.
ViperX trained the first community BS-Roformer model similar to SAMI v1.0 model, although 2 stems (and lower scoring Mel at the time). His BS model sounds similar to Ripple (although it's only 2 stem, while 4 stem Ripple variant scores a bit higher than the 2 stem variant, but still lower than ViperX and v1.0). Then there were a lot of community trained models like private one by ZFTurbo, and fine-tunes by various users (Kim, Unwa, Bas, ZFTurbo)
Bas tried to train a model purely on multisong dataset only, but failed to surpass the SDR score of a 1.1 Bytedance’s model anyway. v1.1 has new arch enhancements to the arch, and will be presented on ISMIR2024 (white paper is already out; link in the Training section).