Ultimate Vocal Remover GUI v4.0.0 (not 4.0.1 for some reason dont ask why) EzoGaming's QOL Patches Edition
Update April 9th 2021: The v5 beta along with 11 beta models have been released! You can read more about it here!
This application is a GUI version of the vocal remover AI created and posted by GitHub user tsurumeso. This version also comes with a total of 11 high performance models trained by anjok7. You can find tsurumeso's original command line version here.
- The Developers
- Anjok07- Model collaborator & UVR developer.
- aufr33 - Model collaborator & fellow UVR developer. This project wouldn't be what it is without your help, thank you for your continued support!
- DilanBoskan - The main UVR GUI developer. Thank you for helping bring the GUI to life! Your hard work and continued support is greatly appreciated.
- tsurumeso - The engineer who authored the original AI code. Thank you for the hard work and dedication you put into the AI code UVR is built on!
- EzoGaming (i swear i changed the username??)
The application was made with Tkinter for cross-platform compatibility, so it should work with Windows, Mac, and Linux systems. However, this application has only been tested on Windows 10. It MIGHT work on Ubuntu, but while the orgiginal was, i don't know if my fork will work or not.
0. Hope to god things work. Python dependencies suck. On my old SSD, i had an install where UVR ran at full speed (24it/s on my 1660 super). And when I installed my new SSD it still ran at that speed, then suddenly i did something (and i dont know what i did) and now i can only get 11.3it/s no matter what i reinstall. So just pray or something.
- Download & install Python 3.7 here (Windows link)
- Note: Ensure the "Add Python 3.7 to PATH" box is checked
- Once Python has installed, download Ultimate Vocal Remover GUI Version 4.0.1 here
- Place the UVR-V4GUI folder contained within the .zip file where ever you wish.
- Your documents folder or home directory is recommended for easy access.
- From the UVR-V4GUI directory, open the Windows Command Prompt and run the following installs -
Personally I just tab to the directory in cmd then activate a conda environment made specifically for this. Then run python vocalremover.py
pip install --no-cache-dir -r requirements.txt
pip install torch==1.8.1+cu111 torchvision==0.9.1+cu111 torchaudio===0.8.1 -f https://download.pytorch.org/whl/torch_stable.html
FFmpeg must be installed and configured in order for the application to be able to process any track that isn't a .wav file, or export as a file other than WAV. Instructions for installing FFmpeg can be found on YouTube, WikiHow, Reddit, GitHub, and many other sources around the web.
- Note: If you are experiencing any errors when attempting to process any media files that are not in the .wav format, please ensure FFmpeg is installed & configured correctly.
- Open the file labeled 'VocalRemover.py'.
- It's recommended that you create a shortcut for the file labeled 'VocalRemover.py' to your desktop for easy access.
- Note: If you are unable to open the 'VocalRemover.py' file, please go to the troubleshooting section below.
- It's recommended that you create a shortcut for the file labeled 'VocalRemover.py' to your desktop for easy access.
- Note: All output audio files will be in the '.wav' format.
- This option allows you to toggle between tsurumeso's v2 & v4 AI engines.
- Note: Each engine comes with it's own set of models.
- Note: The TTA option and the ability to set the N_FFT value is limited to the v4 engine only.
The v2 & v4 AI engines use different sets of models. When selected, the models available for v2 or v4 will automatically populate within the model selection dropdowns.
- Choose Main Model - Here is where you choose the main model to perform a deep vocal removal.
- Each of the models provided were trained on different parameters, though they can convert tracks of all genres.
- Each model differs in the way they process given tracks.
- The 'Model Test Mode' option makes it easier for the user to test different models on given tracks.
- Choose Stacked Model - These models are meant to clean up vocal artifacts from instrumental outputs.
- The stacked models provided are only meant to process instrumental outputs created by a main model.
- Selecting the 'Stack Passes' option will enable you to select a stacked model to run with a main model.
- If you wish to only run a stacked model on a track, make sure the 'Stack Conversion Only' option is checked.
- The wide range of main model/stacked model combinations gives the user more flexibility in discovering what model blend works best for the track(s) they are proessing.
- To reiterate, the 'Model Test Mode' option streamlines the process of testing different main model/stacked model combinations on a given track. More information on this option can be found in the next section.
- The "Codec" option lets you choose between eight common audio codecs. (If.... you count wma to be common... and if people actually use AC3 or Vorbis to encode their music?????)
- Note: The FLAC option is identical in quality to wav, but much smaller, especially for files with more silence.
- Note: Encodes at a bit depth of 16, equivalent to Red Book CD Audio.
- Note: This option, especially the default of FLAC, is very useful for saving space and not having a billion WAV files clogging up the drive. It is also very useful (the Opus option can embed in Discord) for sharing results on Discord without Nitro or fast internet (my speed is only 2mbps, uploading the wavs is a pain, converting them manually is more of a pain)
- Note: I may be biased
- the Bitrate option, as the name suggests, is a slider to control the bitrate of the file. The bitrates cover both Opus and MP3 (6-512, MP3 only supports 8-320). I only included practical values, so i didnt include weird numbers like "278kbps" but did include "208" and "232" for example. If you choose an unsupported bitrate (for MP3), ffmpeg will snap it to the closest one automatically.
- GPU Conversion - Selecting this option ensures the GPU is used to process conversions.
- Note: This option will not work if you don't have a CUDA compatible GPU.
- Nvidia GPUs are the only GPUs compatible with CUDA. If planning on buying a GPU for UVR (or any AI), DO NOT GET AN AMD (Radeon RX Series) GPU or you will be heavily disappointed.
- Note: CPU conversions are much slower compared to those processed through the GPU.
- Note: This option will not work if you don't have a CUDA compatible GPU.
- Post-process - This option can potentially identify leftover instrumental artifacts within the vocal outputs. This option may improve the separation on some songs.
- Note: Having this option selected can potentially have an adverse effect on the conversion process, depending on the track. Because of this, it's only recommended as a last resort.
- TTA - This option performs Test-Time-Augmentation to improve the separation quality.
- Note: Having this selected will increase the time it takes to complete a conversion.
- Note: This option is not compatible with the v2 AI engine.
- Output Image - Selecting this option will include the spectrograms in .jpg format for the instrumental & vocal audio outputs. No clue what this can be used for, I tried messing with them, but there's no way to reinsert them back in, and couldn't figure out how uvr does the spectral stuff
- Model in Filename - A simpler approach to the "Model Test Mode."
- Keep WAV - (Not functional yet) Keep the WAV after processing, if encoding.
- Mono Mixdown - (Not functional yet) Mixdown to mono. Useful if the input is mono and it outputs stereo, or if the input is mono but over two tracks. Also saves space if using a lossless codec, and saves quality if using a lossy codec.
- Stack Passes - This option activates the stacked model conversion process and allows the user to set the number of times a track runs through a stacked model.
- Note: Unless you have the 'Save All Stacked Outputs' option selected, the following outputs will be saved -
- Instrumental generated after the last stack pass
- The vocal track generated by the main model
- Note: The best range is 1-4 passes. 8 or more passes can result in degraded sound quality for the track, and returns diminish severely after 2.
- Note: Unless you have the 'Save All Stacked Outputs' option selected, the following outputs will be saved -
- Stack Conversion Only - Selecting this option allows the user to bypass the main model and run a track through a stacked model only.
- Save All Stacked Outputs - Having this option selected will auto-generate a new folder named after the track being processed to your 'Save to' path. The new folder will contain all of the outputs that were generated after each stack pass. The amount of audio outputs will depend on the number of stack passes chosen.
- Note: Each output audio file will be appended with the number of passes it has had.
- Example: If 5 stack passes are chosen, the application will provide you with all 5 pairs of audio outputs generated after each pass, if this option is enabled.
- This option can be very useful in determining the optimal number of passes needed to clean a track.
- The 'stacked vocal' tracks will contain the audio of the vocal artifacts that were removed from the instrumental.
- These files can be used to verify artifact removal.
- Note: Each output audio file will be appended with the number of passes it has had.
- Model Test Mode - This option makes it easier for users to test the results of different models, and model combinations, by eliminating the hassel of having to manually change the filenames and/or create new folders when processing the same track through multiple models. This option structures the model testing process.
- When 'Model Test Mode' is selected, the application will auto-generate a new folder in the 'Save to' path you have chosen.
- The new auto-generated folder will be named after the model(s) selected.
- The output audio files will be saved to the auto-generated directory.
- The filenames for the instrumental & vocal outputs will have the selected model(s) name(s) appended to them.
- When 'Model Test Mode' is selected, the application will auto-generate a new folder in the 'Save to' path you have chosen.
All models released here will have the values they were trained with appended to the end of their filenames like so, 'MGM-HIGHEND_sr44100_hl512_w512_nf2048.pth'. The '_sr44100_hl512_w512_nf2048' portion automatically sets the SR, HOP LENGNTH, WINDOW SIZE, & N_FFT values within the application. If there are no values appended to the end of a selected model filename, the SR, HOP LENGNTH, WINDOW SIZE, & N_FFT fields will be editable and auto-populate with default values.
(Protip: If on a 6GB GPU, the best training parameters are -sr 33075 -hl 384 -nf 1536)
-
Note - The WINDOW_SIZE value is universal. The smaller your window size, the better your conversions will be. However, a smaller window size means longer conversions times and heavier resource usage. The largest tile size possible on a 6GB GPU is 2048.
- Here are the recommended window size values -
- 1024 - Low conversion quality, shortest conversion time, low resource usage
- 512 - Average conversion quality, average conversion time, normal resource usage
- 320 - Better conversion quality, long conversion time, high resource usage
- 272 - Best conversion quality, longest conversion time, heavy resource usage
- 272 is the lowest window size value possible. (It's a mystery as to why.)
- Here are the recommended window size values -
-
Default Values:
- SR - 44100
- HOP LENGTH - 1024
- WINDOW SIZE - 320
- N_FFT - 2048
- Add New Model - This button will automatically open the models folder.
- Note: If you are adding a new model, make sure to add it accordingly based on the AI engine it was trained on.
- Example: If you wish to add a model trained on the v4 engine, add it to the correct folder located in the 'models/v4/' directory.
- Note: The application will automatically detect any models added the correct directories without needing a restart.
- Note: If you are adding a new model, make sure to add it accordingly based on the AI engine it was trained on.
- Restart Button - If the application hangs for any reason, you can hit the circular arrow button immediately to the right of the 'Start Conversion' button.
All of the models included in the release were trained on large datasets containing diverse sets of music genres.
PLEASE NOTE: Do not change the name of the models provided! The required parameters are specified and appended to the end of the filenames.
Here's a list of the models included within the package -
-
v4 AI Engine
- Main Models
- MGM_MAIN_v4_sr44100_hl512_nf2048.pth - This is the main model that does an excellent job removing vocals from most tracks.
- MGM_LOWEND_A_v4_sr32000_hl512_nf2048.pth - This model focuses a bit more on removing vocals from lower frequencies.
- MGM_LOWEND_B_v4_sr33075_hl384_nf2048.pth - This is also a model that focuses on lower end frequencies, but trained with different parameters.
- MGM_LOWEND_C_v4_sr16000_hl512_nf2048.pth - This is also a model that focuses on lower end frequencies, but trained on a very low sample rate.
- MGM_HIGHEND_v4_sr44100_hl1024_nf2048.pth - This model slightly focuses a bit more on higher end frequencies.
- MODEL_BVKARAOKE_by_aufr33_v4_sr33075_hl384_nf1536.pth - This is a beta model that removes main vocals while leaving background vocals intact. EXTREMELY USEFUL FOR EXTRACTING BACKING VOCALS FROM ACAPELLAS!
- Stacked Models
- StackedMGM_MM_v4_sr44100_hl512_nf2048.pth - This is a strong vocal artifact removal model. This model was made to run with 'MGM_MAIN_v4_sr44100_hl512_nf2048.pth'. However, any combination may yield a desired result.
- StackedMGM_MLA_v4_sr32000_hl512_nf2048.pth - This is a strong vocal artifact removal model. This model was made to run with 'MGM_MAIN_v4_sr44100_hl512_nf2048.pth'. However, any combination may yield a desired result.
- StackedMGM_LL_v4_sr32000_hl512_nf2048.pth - This is a strong vocal artifact removal model. This model was made to run with 'MGM_LOWEND_A_v4_sr32000_hl512_nf2048.pth'. However, any combination may yield a desired result.
- Main Models
-
v2 AI Engine
- Main Models
- Multi_Genre_Model_v2_sr44100_hl1024.pth - This model yields excellent results for most tracks processed through it.
- Stacked Models
- StackedRegA_v2_sr44100_hl1024.pth - This is a standard vocal artifact removal model.
- StackedArg_v2_sr44100_hl1024.pth - This model removes vocal artifacts a bit more aggressively, but may greatly degrade the audio quality of the output audio.
- Main Models
A special thank you to aufr33 for helping me expand the dataset used to train some of these models and for the helpful training tips.
- The application will automatically remember your 'save to' path upon closing and reopening until it's changed.
- Note: The last directory accessed within the application will also be remembered.
- Multiple conversions are supported.
- The ability to drag & drop audio files to convert has also been added.
- Conversion times will greatly depend on your hardware.
- Note: This application will not be friendly to older or budget hardware. Please proceed with caution! Pay attention to your PC and make sure it doesn't overheat. We are not responsible for any hardware damage.
- If using BVKaraoke model, it will change the filenames to match whether you're using a full mix or an acapella.
- This application is not compatible with 32-bit versions of Python. Please make sure your version of Python is 64-bit (it's 2021 anyways. only literal cash registers still use 32 bit software.).
- If FFmpeg is not installed, the application will throw an error if the user attempts to convert a non-WAV file.
Please be as detailed as possible when posting a new issue. Make sure to provide any error outputs and/or screenshots/gif's to give us a clearer understanding of the issue you are experiencing.
If the 'VocalRemover.py' file won't open under any circumstances and all other resources have been exhausted, please do the following -
- Open the cmd prompt from the UVR-V4GUI directory
- Run the following command -
python VocalRemover.py
- Copy and paste the error output shown in the cmd prompt to the issues center on the GitHub repository.
The Ultimate Vocal Remover GUI code is MIT-licensed.
- PLEASE NOTE: For all third party application developers who wish to use our models, please honor the MIT-license by providing credit to UVR and it's developers Anjok07, aufr33, & tsurumeso.
- For anyone interested in the ongoing development of Ultimate Vocal Remover GUI please send us a pull request and we will review it. This project is 100% open-source and free for anyone to use and/or modify as they wish.
- Please note that we do not maintain or directly support any of tsurumesos AI application code. We only maintain the development and support for the Ultimate Vocal Remover GUI and the models provided.
- [1] Takahashi et al., "Multi-scale Multi-band DenseNets for Audio Source Separation", https://arxiv.org/pdf/1706.09588.pdf