From 575bbf3c998d9128d40805a2db0d3f7255f13549 Mon Sep 17 00:00:00 2001 From: Anjok07 <68268275+Anjok07@users.noreply.github.com> Date: Thu, 7 Apr 2022 02:13:52 -0500 Subject: [PATCH] Update README.md --- README.md | 122 +++++++++++++++++------------------------------------- 1 file changed, 38 insertions(+), 84 deletions(-) diff --git a/README.md b/README.md index 0bf5373..c0cab44 100644 --- a/README.md +++ b/README.md @@ -1,4 +1,4 @@ -# Ultimate Vocal Remover GUI v4.0.1 +# Ultimate Vocal Remover GUI v5.0.0 [![Release](https://img.shields.io/github/release/anjok07/ultimatevocalremovergui.svg)](https://github.com/anjok07/ultimatevocalremovergui/releases/latest) @@ -6,14 +6,12 @@ ## About -**Update April 9th 2021:** The v5 beta along with 11 beta models have been released! You can read more about it [here](https://github.com/Anjok07/ultimatevocalremovergui/tree/v5-beta-cml)! - -This application is a GUI version of the vocal remover AI created and posted by GitHub user [tsurumeso](https://github.com/tsurumeso). This version also comes with a total of 11 high performance models trained by me. You can find tsurumeso's original command line version [here](https://github.com/tsurumeso/vocal-remover). +This application is a GUI version of the vocal remover AI created and posted by GitHub user [tsurumeso](https://github.com/tsurumeso). This version also comes with eight high-performance models trained by me. You can find tsurumeso's original command-line version [here](https://github.com/tsurumeso/vocal-remover). - **The Developers** - [Anjok07](https://github.com/anjok07)- Model collaborator & UVR developer. - - [aufr33](https://github.com/aufr33) - Model collaborator & fellow UVR developer. This project wouldn't be what it is without your help, thank you for your continued support! - - [DilanBoskan](https://github.com/DilanBoskan) - The main UVR GUI developer. Thank you for helping bring the GUI to life! Your hard work and continued support is greatly appreciated. + - [aufr33](https://github.com/aufr33) - Model collaborator & fellow UVR developer. This project wouldn't be what it is without your help. Thank you for your continued support! + - [DilanBoskan](https://github.com/DilanBoskan) - The main UVR GUI developer. Thank you for helping bring the GUI to life! Your hard work and continued support are greatly appreciated. - [tsurumeso](https://github.com/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! ## Installation @@ -24,10 +22,10 @@ The application was made with Tkinter for cross-platform compatibility, so it sh 1. Download & install Python 3.7 [here](https://www.python.org/ftp/python/3.7.0/python-3.7.0-amd64.exe) (Windows link) - **Note:** Ensure the *"Add Python 3.7 to PATH"* box is checked -2. Once Python has installed, download **Ultimate Vocal Remover GUI Version 4.0.1** [here](https://github.com/Anjok07/ultimatevocalremovergui/releases/download/v4.0.1/UVR_V4GUI_All_IN_ONE_12_10.zip) -3. Place the UVR-V4GUI folder contained within the *.zip* file where ever you wish. +2. Once Python has been installed, download **Ultimate Vocal Remover GUI Version 5.0** [here](Link-coming) +3. Place the UVR-V5GUI folder within the *.zip* file where you wish. - Your documents folder or home directory is recommended for easy access. -4. From the UVR-V4GUI directory, open the Windows Command Prompt and run the following installs - +4. From the UVR-V5GUI directory, open the Windows Command Prompt and run the following installs - ``` pip install --no-cache-dir -r requirements.txt @@ -36,9 +34,9 @@ pip install torch==1.8.1+cu111 torchvision==0.9.1+cu111 torchaudio===0.8.1 -f ht ### FFmpeg -FFmpeg must be installed and configured in order for the application to be able to process any track that isn't a *.wav* file. Instructions for installing FFmpeg can be found on YouTube, WikiHow, Reddit, GitHub, and many other sources around the web. +FFmpeg must be installed and configured for the application to process any track that isn't a *.wav* file. 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. +- **Note:** If you are experiencing any errors when attempting to process any media files, not in the *.wav* format, please ensure FFmpeg is installed & configured correctly. ### Running the Vocal Remover GUI & Models @@ -48,112 +46,68 @@ FFmpeg must be installed and configured in order for the application to be able - **Note:** All output audio files will be in the *'.wav'* format. ## Option Guide - -### Choose AI Engine: - -- 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. ### Model Selections: -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'*](https://github.com/Anjok07/ultimatevocalremovergui/tree/beta#checkboxes) 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'*](https://github.com/Anjok07/ultimatevocalremovergui/tree/beta#checkboxes) 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'*](https://github.com/Anjok07/ultimatevocalremovergui/tree/beta#checkboxes) 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'*](https://github.com/Anjok07/ultimatevocalremovergui/tree/beta#checkboxes) 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. ### Checkboxes - **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. - - Nividia GPU's are most compatible with Cuda. - - **Note:** CPU conversions are much slower compared to those processed through the 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. + - Nvidia GPU's are most compatible with Cuda. + - **Note:** CPU conversions are much slower than those processed through the GPU. +- **Post-process** - This option can potentially identify leftover instrumental artifacts within the vocal outputs. This option may improve the separation of *some* songs. + - **Note:** Having this option selected can adversely affect 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. -- **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 3-7 passes. 8 or more passes can result in degraded sound quality for the track. -- **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. -- **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. +- **Model Test Mode** - This option makes it easier for users to test the results of different models and model combinations by eliminating the hassle of manually changing the filenames and/or creating 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. + - The filenames for the instrumental & vocal outputs will have the selected model(s) name(s) appended. ### Parameter Values -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. - -- **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. - - Here are the recommended window size values - +- **Window Size** - The smaller your window size, the better your conversions will be. However, a smaller window means longer conversion times and heavier resource usage. + - Here are the selectable 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.* + +- **Aggressiveness Setting** - This option allows you to set how strong the vocal removal will be. The range is 0-100. The higher the value, the more the vocal data will be removed. Higher values can result in muddy-sounding instrumentals for instrumental models depending on the track being converted, so this isn't always recommended. The default is 10 for instrumental & vocal models. - **Default Values:** - - **SR** - 44100 - - **HOP LENGTH** - 1024 - - **WINDOW SIZE** - 320 - - **N_FFT** - 2048 + - **Window Size** - 320 + - **Aggressivness Setting** - 10 ### Other Buttons: -- **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. -- **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. +- **Open Export Directory** - This button will open your 'save to' directory. +- **Restart Button** - If the application hangs for any reason, you can hit the circular arrow button to the right of the *'Start Conversion'* button. ## Models Included 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. + - HP2 - These models have more capacity but are resource heavy! + - HP - These models use the standard UVR v5 model capacity. 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. - - **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. - -- **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. - -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. +- **Main Models** + - **HP2_3BAND_44100_MSB2.pth** - This is a strong instrumental model trained using more data and new parameters. + - **HP2_4BAND_44100_1.pth** - This is a strong instrumental model, but also resource heavy. + - **HP2_4BAND_44100_2.pth** - This is a fine tuned version of the HP2_4BAND_44100_1.pth model. + - **HP_4BAND_44100_A.pth** - This is a strong instrumental model. + - **HP_4BAND_44100_B.pth** - This is a fine tuned version of the HP_4BAND_44100_A.pth model. + - **HP_KAROKEE_4BAND_44100_SN.pth** - This is a model that removes main vocals while leaving background vocals intact. + - **HP_Vocal_4BAND_44100.pth** - This model emphasizesvocal extraction. The vocal stem will be clean, but the instrumental might sound muddy. + - **HP_Vocal_AGG_4BAND_44100.pth** - This model also emphasizes vocal extraction and is a bit more aggressive than the previous model. ## Other GUI Notes @@ -161,8 +115,8 @@ A special thank you to aufr33 for helping me expand the dataset used to train so - **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.*** +- Conversion times will significantly depend on your hardware. + - **Note:** This application will *not* be friendly to older or budget hardware. Please proceed with caution! Please pay attention to your PC and make sure it doesn't overheat. ***We are not responsible for any hardware damage.*** ## Troubleshooting @@ -188,11 +142,11 @@ python VocalRemover.py The **Ultimate Vocal Remover GUI** code is [MIT-licensed](LICENSE). -- **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. +- **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 its developers Anjok07, aufr33, & tsurumeso. ## Contributing -- 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. +- 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. ## References