I've just been sticking with Google Colab until a final is released or the conversion process is updated in Google Colab. @djtayz may leave instructions for a "fileless method in Google Colab. Waiting for that one as well.
Printable View
I've just been sticking with Google Colab until a final is released or the conversion process is updated in Google Colab. @djtayz may leave instructions for a "fileless method in Google Colab. Waiting for that one as well.
***UPDATE***
Here's the beta release for the Ultimate Vocal Remover GUI v4 with 3 brand new models!
Install Instructions:
1. Install Python via the following link and make sure to check the box that says "Add Python 3.6 to PATH" - https://www.python.org/ftp/python/3.....6.8-amd64.exe
2. Open the cmd prompt and run the following -
pip install Pillow
pip install tqdm==4.30.0
pip install librosa==0.6.3
pip install opencv-python
pip install numba==0.48.0
pip install SoundFile
pip install soundstretch
pip install torch==1.6.0+cu101 torchvision==0.7.0+cu101 -f https://download.pytorch.org/whl/torch_stable.html
3. Download Ultimate Vocal Remover GUI v4 Beta here - http://www.mediafire.com/file/nybrcj...-BETA.zip/file
4. Extract the V4GUI-BETA folder to your documents from the zip file
5. Now go into the V4GUI-BETA folder and double click the VocalRemover_v4.py to verify it works
6. Create and place a shortcut for the VocalRemover_v4.py file to your desktop for easy access
7. *****READ EVERYTHING BELOW FOR PROPER CONVERSIONS!*****
Here's a list of the models (THE SR & HOP LENGTH VALUES MUST BE IN LINE WITH THE MODEL IN ORDER FOR YOUR CONVERSIONS TO COME OUT RIGHT!):
- MGM-32000-512.pth - Set the SR to 32000 and the HOP LENGTH to 512 before doing conversions with this model!
- MGM-44100-512.pth - Set the SR to 44100 and the HOP LENGTH to 512 before doing conversions with this model!
- MGM-44100-1024.pth - Set the SR to 44100 and the HOP LENGTH to 1024 before doing conversions with this model!
A few notes & added features:
- Remembers "Save to" and last folders accessed.
- The "Add New Model(s)" button automatically opens the models directory. When you add new models to the appropriate folders, the application will automatically detect them so you don't have to restart it.
- I only have instrumental models for v4 at this time. I still need to train a few stacked models and a vocal model so ignore the options for those... for now.
- Keep in mind this is the beta so please feel free to report any bugs to me here!
- More enhamcements will be made as well
***TROUBLESHOOTING***
- If the VocalRemover_v4.py file won't open under any circumstances, please do the following
1. Open the cmd prompt from the V4GUI-BETA directory
2. Run the following - python VocalRemover_v4.py
3. Copy and paste the error in the cmd prompt here for further assistance
Link:
http://www.mediafire.com/file/nybrcj...-BETA.zip/file
I think the biggest issue will be remembering to go back and point or set ffmpeg for Vocal Remover that we had to do to start with.
OK, so I've been playing with the new beta release just posted and I'm finding that the vocal removal isn't quite on par with the current Google Colab results. I tried all 3 models and it appears that for the song I already have done using the Google Colab method, I don't really hear any improvement differences. That's not saying it's useless. It's saying I will have to mess with another track to hear for sure that Google Colab can't currently do yet. It will probably have to be when the final release is finished.
- MGM-32000-512.pth - This model is very good at capturing lower end frequencies. So, on tracks that convert poorly on all of the other models should come out well on this one.
- MGM-44100-512.pth - My tests of this model have shown this one to actually be the best one I've done to date.
If you used the GoogleColab version, what command did you use to do the inference?
They are great sounding models, though I have not had a ton of time to test them yet. Life stuff is finally settling down now, so I just gotta wait for my headphones to get back from the shop so I can listen to the conversion results as best as possible.
Curiously though, (at least through my speakers), I felt like 32000-512 did the best job of the three on Rush - The Pass, though it was somewhat close with 44100-512 and 44100-1024 did the worst by a real margin. Geddy is a very high singer, so I found this interesting. I'm looking forward to trying some low range stuff though once my headphones are fixed :)
On Google Colab I usually use the multi model and the NP model for a comparison for most tracks. on the GUI I do run all 3 models to see which sounds best and usually Google Colab always comes out better for some reason. Or Google and GUI will come out the same in some cases. I'm not complaining as I'm sure due to single mixed tracks it'll never be 100% like the studio. But for the most part Google Colab has helped me in my needs the best. For anyone wanting cleaner stems run a track through the vocal remover first, then run the instrumental through demucs. I find the stems much cleaner doing that.
@NewAgeRipper - I can't respond to your message. Says you exceeded your storage.
***GUI BETA UPDATE***
================================================== =========================
Bug Fixes -
~The application no longer cuts name off of some filenames after conversions.
~ Application now accepts all file types compatible with ffmpeg
~ (please install ffmpeg prior to running anything other than a wav file)
Changes -
~ The application will now read model parameters from filename (if present)
~ For example, a model with the filename "MGM-LOWEND_sr32000_hl512_w512_nf2048" will automatically fill the SR, HOP LENGTH, WINDOW SIZE, & N_FFT values
~ If the filename was "MGM-LOWEND", the SR, HOP LENGTH, WINDOW SIZE, & N_FFT values will auto-populate with the defaults.
~ The application reads these values from the following portion of the file "_sr32000_hl512_w512_nf2048"
~ A new option called "Model Test Mode" has been added.
~ This option is meant to make it easier for users to test the results of different models without having to manually create new folders and/or change the filenames.
~ When it's selected, the application will automatically generate a new folder with the name of the selected model in the "Save to" path you have chosen.
~ The completed files will have the selected model name appended to it and be saved to the auto-generated folder.
Here's a list of the models (PLEASE DO NOT CHANGE THE NAME OF THE FIRST 2 MODELS LISTED AS THE PARAMETERS ARE SPECIFIED IN THE FILENAMES!):
- MGM-LOWEND_sr32000_hl512_w512_nf2048.pth - This model is good at capturing vocals on the low end of the spectrogram.
- MGM-44100-512_sr44100_hl512_w512_nf2048.pth - This is a multi-genre model that was trained with a hop length size of 512. It's debatably the best model of this group.
- MGM-44100-1024.pth - This is a multi-genre model trained with basic parameters.
Link: http://www.mediafire.com/file/q5xefq...v1102.zip/file
================================================== =========================
Here's the beta release for the Ultimate Vocal Remover GUI v4 with 3 brand new models!
Install Instructions:
1. Install Python via the following link and make sure to check the box that says "Add Python 3.6 to PATH" - https://www.python.org/ftp/python/3.....6.8-amd64.exe
2. Open the cmd prompt and run the following -
pip install Pillow
pip install tqdm==4.30.0
pip install librosa==0.6.3
pip install opencv-python
pip install numba==0.48.0
pip install SoundFile
pip install soundstretch
pip install torch==1.6.0+cu101 torchvision==0.7.0+cu101 -f https://download.pytorch.org/whl/torch_stable.html
3. Download Ultimate Vocal Remover GUI v4 Beta here - http://www.mediafire.com/file/nybrcj...-BETA.zip/file
4. Extract the V4GUI-BETA folder to your documents from the zip file
5. Now go into the V4GUI-BETA folder and double click the VocalRemover_v4.py to verify it works
6. Create and place a shortcut for the VocalRemover_v4.py file to your desktop for easy access
***TROUBLESHOOTING***
- If the VocalRemover_v4.py file won't open under any circumstances, please do the following
1. Open the cmd prompt from the V4GUI-BETA directory
2. Run the following - python VocalRemover_v4.py
3. Copy and paste the error in the cmd prompt to the technical channel for further assistance
will you update it on Colab Google soon?
We're still trying to determine if there is really much difference in improvement to warrant updating Google Colab. I need @Anjok to tell me the songs he's tested that make the 44100-512 model better than previously to compare between the 2. Also no one else has been giving any feedback on this really.
I have noticed a "bug" with short songs. Not just songs, any audio file in general, that is less than 30-40 sec. (haven't tested that).
The conversion fails at the end of the sample, essentially not removing the vocals from the song, kinda weird.
Here's a recent sample I did, but I noticed that this has been happening for a long time https://wetransfer.com/downloads/b5e...5100341/2e003b
I'm still waiting for my headphones. Can't really test anything until I get them back. My speakers are pretty good, but it's hard to hear the small differences on tracks that convert well. Need the headphones for that.
@djtayz believe it or not you can actually make an instrumental, then invert it to the original song you used and get a better-ish pella than either of the colabs give. Thought you might wanna play with that.
I just tried it but the results are the same (as expected I think), unless you mean something different.
Combining Vocal Remover's instrumental and acapella tracks, then inverting it with the original should give you silence, that means no loss in quality
If you're referring to my last post (about the "bug"), I just wanted to say that when using shorter audio files the instrumental track adds back the vocals at the end, as supposed to having them removed completely, which I find odd
@djtayz has there been any new updates to the Google Colab yet?
@Anjok, would it be possible to make the GUI resizable?
I'm personally not able to run the newest models that Anjok released with the most recent GUI, another version of Colab is in development though (not by me).
I will see if I can fix things, meanwhile, I'm open to any kind of requests for Colab, will try my best to implement them or think of workarounds.
Hi everyone.
I have a problem. I'm using this great GUI for months. To day I installed demucs using this guide https://github.com/facebookresearch/demucs
I didn't install python 3.7 because it is already installed. I use demucs and I think it it great too for some side but now when I try to open vocal remover the prompt appears just for a second then disappears and the gui doen's run. What happened? What can I do? I'd like to use both... please help! Thx
Read about the beta update HERE or read the instructions on the github as linked in the main TOC.
Another question: Everytime I use google colab I have to do all the steps from the beginning? Or there 's a way to save the driver?
The small .ipynb file should be fine always after uploading once. But all the other steps have to be done each time. Only if you experience an error converting a song that I would reup the .ipynb file to start it as new again. But the big folder you upload to your google drive is fine always and shouldn't need to be redone again. Hope this helps. You can always from Google Colab go to "File" "Open Notebook" and remove the previous sessions and browse for the .ipynb file from your hard drive again to start as new. I honestly need to update the video because at that time I thought you had to upload the .ipynb file every time. As I said it doesn't hurt to help avoid corruption in Colab.
The .ipynb file is what you do the first time you opened Google Colab to upload it. after that Google Colab sees it in your drive. That is why Step 1 requires authentication, step 2 checks for ffmpeg items. Step 3 checks and redownloadeds within the cloud to satisfy everything else. Step 4 looks for synced vocal remover folder. Then you choose step 5 for your choice of models to use.
As I said it's an amazing tool with great models, but it would be awesome if the Google Colab version could be updated because due to my equipment, it's taking a lot of time to process and I have to be really patient lol
Thanks in advance...
Ah sorry, I was talking about the GUI ^^
You'll soon learn that Google Colab and the GUI tool actually goes hand in hand depending on your needs. I find that the low end 32000-512 model is best for most tracks that I need the vocals out of. It seems to smooth out the artifacts from certain vocal pitches better. I render one in the GUI in 22 to 40 min depending on the song. I'm a karaoke enthusiast so if the slower method gives me the best results, then that is what I go with. Also the GUI is handy for when you can't be on an internet connection such as a laptop or Windows tablet. Im working with songs that never made it into karaoke labels and even know how to separate lead and backing vocals. The more you mess with it the more ideas you get.