Been using a variety of Voice AI projects this far - SO-VITS-SVC, DDSP-SVC and also RVC - All give different results obviously, SO-VITS retains the voice characteristics (e.g. accents) a bit more than RVC and DDSP, but RVC has the best overall resulting audio But the fact we can get these results already? Absolutely crazy, can't wait for what the v3 model is gonna bring to the table! :) Also thank you so much for doing all these videos! I always find some new stuff :D
RVC is actually pretty dang amazing. Already trained and combined voices. Input audio really does matter. Getting super clean audio that has never been compressed is the only way to go for best results. Also pitch shift the target audio that will be getting replaced to match the pitch of the training voice data _using an outside program_ any good DAW will do. Much much better results. The built in formant shifter is not very good and very quickly creates bad distortion with large pitch changes. It works better for transforming to higher pitches but doing it in a DAW is still the way to go as you can get the pitch matching to be pretty close to exact. Even with how reasonably it can work with any voice I find that this seems to lend itself to turning your voice into anime girls to have by far the cleanest results -- Least roboty. Like it was literally made for it. Built in UVR5 voice isolation to rip voice out of music to make acapellas for training set. It's not perfect. Trial and error here works best when the melodies between voice and background are not the same. Overlapping melodies will lead to bleed through. Now all RVC needs is AMD ROCm support on Windows (similar to Nvida CUDA for ML). Trying to get it working through Windows Linux subsystem (ROCm support from this software is Linux only for now) but with little success. Still says no GPU found. Could very well be my being a Linux noob. Training on larger datasets (couple GB) with CPU only is bonkers slow but it works well. Think of it like this. If your training data is an hour long, on CPU assuming it's appreciably beefy, it will probably take about an hour per epoch. Obviously with GPU this is much much faster if you have the VRAM for it. I am using RVC1006AMD_Intel Git. I have used different versions which are faster in CPU training. Hope they update the package soon with proper AMD GPU support. Not all forks are equal. My antivirus goes crazy on some of the custom forks. The go-web.bat file not editable. That's a red flag. It could be false positive but nevertheless be careful of bad actors. The OG author's builds are editable and don't raise any red flags. All of the source code is readily available for you to look at and probe for any external connections which could be hidden. I did not find anything fishy in the OG's builds after scanning every .bat and .py file with my eyeballs.
Thank you again! This tutorial has more details than the last one you made. A lot more is explained and I really appreciate this tutorial that you have made.
aweseome stuff. i will definitely try it out. my hardest challenge would be to restore the voice from my gradparents etc from old restored reel2reel tapes... i barerely get 5 minutes or so in medium quality. lots of work, but definitely worth it. thx for bringing this up and explaing this stuff in very detail. i absolutely love your content. i already made my grand.grand.mothers picture saying something, which lead into a spontaneous flood of tears ... :))
I realise that You do not have to sing for the training DATA, but I did and felt a right prat haha. I live in a block of flats with neighbours above and below. Thankfully I have a spare room that's pretty well loaded with boxes etc, so not too reverb(y) Anyway, I got a nice clean recording of me singing about how I'm singing solely for the purpose of creating training DATA. And other mundane things like apologising to the neighbours, that it would only be for 10 minutes etc 🙂 I used my trusty Tascam H4n Pro, and wind shield My PC is just churning away now, and I'm looking forward to experimenting with this. Thanks Nerdy for reminding me again about this project.
Unfortunately when you run the AMD/Intel version you encounter the following error: "NO GPU DETECTED: falling back to CPU - this may take a while". running on your CPU is extremely slow and will take days. Please gives us an update if you find a way to make this actually work for those of us with with AMD GPUs.
Nice work! So let me get this straight...I could clone my own voice reading a paragraph with specific tones, intonations, etc...and then I could clone a second voice and have that second voice read back my original recorded paragraph with the exact same specific tones, intonations, etc...?
Love your stuff - searching the internet for the best AI cover apps to use to share later. Thanks for your in-depth vid. You got a new sub and a like for your spectacular input.
An error in cmd shows like that when i trying to convert if data.dtype in [np.float64, np.float32, np.float16]: AttributeError: 'NoneType' object has no attribute 'dtype' why this happens please help
neat! as always, this is beyond me... :D but I'll hand it over to my brother who might figure it out. Lets see Weird Al doing serous songs! or whatever random things!
I tried just installing and unzipping and starting go-web, but it doesn't exactly work. The CMD starts, but it can't find the specified path and that's where it can't go further. I'm no coder and I don't understand python, but I would like to know what I'm missing since this is clearly not install, unzip, and have fun.
@@NerdyRodent Looks like I installed everything right, just not in the exact way you did. For instance, I extracted with winrar instead of 7zip the first time, and I downloaded the correct file, but not from the exact same place. Just followed the exact steps you did and it worked this time. Thanks for the response!
have the same problem- the huggingface page also dont have any zipfile in the directory anymore. i downloaded it from github, but "it can't find the specified path". is it dead?
wow ! Is installation in windows via command prompt similar to the linux method ? the usual pip pytorch etc etc Also, is anaconda necessary to the process or can we make our own virtual environments. Thanks a bunch mate
thank you for the video, very good explanation. i have a question, my training seems to be doing about 4:30min per epoch on my 1080ti at 10-11 batch size (any higher doesnt seem to work) and like 50% GPU usage. is this what i should expect from this card? do you happen to know any good cloud solutions? thanks again
Sounds about right for a card of that age. Cloud options will depend a little of where you are in the world, but AWS, paperspace and others offer GPU systems
Hello! First video I've seen from you, thank you very much for this excellent tutorial, helped me out a lot! I've got a question though: Is the webinterface available from outside of my machine? Can other people access it? It seems as if I am hosting it, and the batchfile says "go-web", so I wonder if there is any security concern? Cheers, looking forward to more wonderful content from you!
There is a docker version, I'm definitely going to try that, why go through the pain of installing pytorch etc when you can just git pull and docker compose up --build?
@@NerdyRodent I second that 100%! I just have too much going on on my PC and I'm always scared of these things interfering with each other, that's why I love containers
so how do i use/setup tensorboard to check if im overtraining and when and how should i stop the training of a voice cause i have no clue what im doing
Do training files have to be in english or can you throw any language at it? And when you extract voice from a music clip, can you also take any language?
Been using RVC for a couple weeks and the accents aren't as strong , voices kind of mimic the accents of the original singer even when set to max.. any solution?
Fab tutorial as always! Thanks 😁 Btw I tried following your previous RVC tutorial and I couldn’t get good sounding results like in your demo. Not really sure what I did wrong, maybe a poor dataset but the voice was all screechy and distorted and robotic. Will try again!
i was wondering if you would create a tutorial on the txt2speech repo of MRQ ai voice cloning, its toriose doesn't take much time to train and the output is decent and theirs a lot of settings to play around with.. though not knowledgable and how to get the best results, it would be nice if you would go into that in ur tutorial if it felt necessary, but its a cool workflow to go from MRQ txt2speech to RVC speech2speech to get enhanced results, it brings things much closer to 11labs quality... but not in convenience for now
Can we do voice cloning using this above method and previously i was trying with a google collab project and i got the some runtime error. Does this project includes any google collab and stuffs clarify this. Thanks in advance.
There was a colab, but this isn’t really what colab is for. If you’ve not got your own computer to use, there are numerous rental options such as paperspace, AWS, etc.
@@NerdyRodent i.imgur.com/vWnkX8k.png Was this page i oddly couldn't find. Which looks to be models trained by others as you noted :) Edit, found it out, just search for "rvc", and they all pop up xD. Sorry for wasting a bit of time, just takes a click or 2 before i get "basics" :P
Hey, Nerdy Roden! How is this differnt than Bark AI you did a tutorial a while ago? Is this better? Also can it be used for something else besides songs? Narration for example?
Not had any issues myself, but the full error message should indicate what went wrong. Make sure to use the correct requirements file for your hardware.
I can't figure out how to get this to run on my Mac M1. Is there a very basic step by step of how you run this? I have no knowledge of python or any of this environment
I am not having any success with the Visions of Chaos build of this on my PC though your video is the first I have seen that explains the GUI and confirms I am using that correctly at least so that's not the issue, no failures just ...nothing, it never completes the first step
just have a question. How high is your batch size, when you train? Is it something that if you set it too high, you get an imprecise module? If I have a dataset of one hour, what should my batch size be?
@Nerdy Rodent, I have extracted the vocals of a pretty heavy death metal song. The vocals are scuffed a bit but thats fine. My issue is that the instrucmentals have alot of static, i used HP3, should i just try the others?
@@NerdyRodent Oh yeah, that totally worked better. There was still some echo or reverb (not sure). Honestly what ive first created is really cursed because Mr Krabz doesnt like harsh vocals but i think it makes it even better for a first attempt... Anyway ill try to tweak tomorrow to see how i can make it better
Hello sir! Great video :) But I need help. Every time I try to clone the voice, an error occurs: "nonetype' object has no attribute 'dtype'". And I don't know why. I name the file well (including the .wav), but it always marks error. The weirdest is that some few files work well, but most of files don't work
Thanks! This is almost what I need. However I want to replace the original lyrics with my own and vocalized by f.i. Elvis. Any chance of this happening with this software or some other?
Zip file first, anaconda and git second. Either method should download all files correctly. Other things that could cause files to not download would be things like faulty anti-virus products, faulty drive, etc.
in the new version of RVC, which has just arrived you have to put your modules inside assets/whaights. This is a problem when I have now made zip files that I have repackaged, so log and wheights are in the same zip file, so no assets folder. Can you change it? Now it detects itself which index file belongs to the module that you have now loaded
I'm on an Intel Mac. Been working at this trying to get this going for a WHILE. Finally got it running, but now it's seeming like I need a GPU in order for it to work on my computer. In the script one of the last things it says is "move model to cpu". How do I do this?
I keep getting "Connection errored out" after a few seconds of doing anything in the web UI. I haven't closed the command prompt, and gave exceptions for the antivirus. What could it be?
Hi. I have a background in audio production & want to ask a existential question. Creatives minds are using these tools in positive ways but what about the negatives. . Do you think it's possible to clone a voice so the results will null when compared with the original source material. Will voice verification tools be mandatory in the future?
There's this song written by my favourite band (A) that was sold and sang by a different vocalist (B). What I want to do is extract the vocals of (B), create a voice model of (A), and replace (B) with (A), so that the original writer actually sings this song. Is this possible?
@@NerdyRodent Same for me , I never trusted it ! There was always a risk someone can make the colab stop working and there is nothing you can do about it. That is the reason why I quickly preferred to install the RVC GUI for MAC and it works so perfect for inferencing ! However I wasn't successful to get the training part to work correct. I am still searching some support how to do this but everyone seems to work on PC. Now that the colab training has gone I will need to searcher harder how to get the MAC training done for RVC GUI, but thanks for reply !
Thanks for the video i have a 4090 and noticed with a similar sized data set if i set the batch size to 40 it was about 5 seconds a epoch but if i set it to a multiple of 8 like 32 it was giving me epochs in 2-3 seconds. Also the readme mentions access to v3 but the ui only showed v1 and v2 as options? Have you tried the real time voice changer they link by Okada?
man, so i used the google colab, rvc v2 thing for a while now, and it was great, but now it just disconnects when getting to training step, this does not have the same problem? Do you know a way around this problem?
Hello, I cannot train a voice, the site does not recognize the address where the file is located, I tried everything, changed folders, .mp3, .wav, .pth, nothing there do ! Especially since I already have the .pth file, can you help me please?
Think i have to wait for the author patches it out. Using stock voices, and hit convert, and i get this AttributeError: 'NoneType' object has no attribute 'dtype' And error above saying " ffmpeg._run.Error: ffmpeg error (see stderr output for detail) " And Can't seem to find this stderr log.
@@NerdyRodent Aye. Got it all working now, just took miniconda and made a venv instead :) And tested a 10 min audio sample, and 200 epochs was not enough, so doubling it saving every 20 epochs until i find the best ratio :P
@@NerdyRodent Yep! :D By the way, do you know of a better text to voice than bark that can read pre trained voices in .pth format like the one in this video produces? As bark also uses seeds which is kind of annoying as i'm lacking a module to set a manual seed, and lacks "use last seed", and it's results is quite hit and miss.
Thank you for this video! Your video was very helpful, but when I use the female voices, they don't sound at all like female voices; instead, they sound robotic-like masculine voices. How do I make them work properly?
Oh, I think I figured out the issue, and the female voices now work fine. All I had to do was change the transposition value to around 10-12 instead of the default value of 0. Regardless, thank you for your quick response!@@NerdyRodent
Hey, trying to separate the vocals and instrumental using the GUI, but I'm getting an error every time despite following the same steps as you. I could paste the output information here, but it's quite lengthy. Think you could help out?
I'm really not understanding your guide for the portable version, I don't have the asset folder when I download the zip, and even in your video you don't have it either then you say to jump to the next section and you all the sudden have that folder, am I suppose to download that folder separately or what?
@@NerdyRodent i appreciate the response, but are you saying the portable version isn't updated? or the older models don't have an asset directory? Am i suppose to do something different in order to use a v2 model? Sorry if this is basic knowledge, im just having a little difficult time understanding it
this is the daft punk vocoder model. How can i train this? altso with out v2 Output information Model information: 200epoch Sampling rate: 32k Whether the model inputs pitch guidance: 1 Version: None
Yeee. I'm like, is he joking? It's one of the top languages. Enjoy took from various languages, too. I'm part French. Hehe. :3 22% in DNA. I remember the Dexter's Laboratory episode of French speaking back in the days when it came out. Hehe :3
@@NerdyRodent I’m very new to this , I just bought my stationary setup. I tried right clicking in the 7zip but it didn’t show up - as for operating system , I’m not sure , I have windows 11 though .
Windows 11 (assuming you’re fully up to date) should be able to unzip 7zip files ok as of Build 23493. If you have an older build, then you will need to download the 7zip program as Microsoft Windows is too out of date to support modern files, such as 7zip. Go to www.7-zip.org/ to download a special program to make windows able to unzip files.
Thats great! Also I am not using any pitching correction I just do all that in audacity before converting, so if I do that, then is rmvpe still better quality than crepe and harvest? And thanks for the reply and awesome video tutorial, cheers!
the portable version did not work for me. Windows 10, Nvidia 3080. error importing numpy "Importing the numpy C-extensions failed. Original error was: DLL load failed while importing _multiarray_umath: The specified module could not be found." I posted it on github, but in case somebody knows the solution for the portable version, please let me know...
Some day can you please make a dedicated video on how to install multiple AI tools and save Precious HDD/SDD space too. At this point we have many AI tools, each requiring different packages but the same python version and for some AI tools a different version and combination of requirements. Is there a way to avoid multiple installations of the same GBs worth of packages and Just point to the existing under Linux?
Been using a variety of Voice AI projects this far - SO-VITS-SVC, DDSP-SVC and also RVC - All give different results obviously, SO-VITS retains the voice characteristics (e.g. accents) a bit more than RVC and DDSP, but RVC has the best overall resulting audio
But the fact we can get these results already? Absolutely crazy, can't wait for what the v3 model is gonna bring to the table! :)
Also thank you so much for doing all these videos! I always find some new stuff :D
V3 is gonna be awesome! (Crosses fingers)
vall-e-x is good also with 3 seconds of audio they claim i just installed it
It’s just a shame vall-e-x clones sound absolutely nothing like the voice you’re trying to actually clone… nor can they sing 🙁
RVC is actually pretty dang amazing. Already trained and combined voices. Input audio really does matter. Getting super clean audio that has never been compressed is the only way to go for best results. Also pitch shift the target audio that will be getting replaced to match the pitch of the training voice data _using an outside program_ any good DAW will do. Much much better results. The built in formant shifter is not very good and very quickly creates bad distortion with large pitch changes. It works better for transforming to higher pitches but doing it in a DAW is still the way to go as you can get the pitch matching to be pretty close to exact. Even with how reasonably it can work with any voice I find that this seems to lend itself to turning your voice into anime girls to have by far the cleanest results -- Least roboty. Like it was literally made for it.
Built in UVR5 voice isolation to rip voice out of music to make acapellas for training set. It's not perfect. Trial and error here works best when the melodies between voice and background are not the same. Overlapping melodies will lead to bleed through.
Now all RVC needs is AMD ROCm support on Windows (similar to Nvida CUDA for ML). Trying to get it working through Windows Linux subsystem (ROCm support from this software is Linux only for now) but with little success. Still says no GPU found. Could very well be my being a Linux noob. Training on larger datasets (couple GB) with CPU only is bonkers slow but it works well. Think of it like this. If your training data is an hour long, on CPU assuming it's appreciably beefy, it will probably take about an hour per epoch. Obviously with GPU this is much much faster if you have the VRAM for it. I am using RVC1006AMD_Intel Git. I have used different versions which are faster in CPU training. Hope they update the package soon with proper AMD GPU support.
Not all forks are equal. My antivirus goes crazy on some of the custom forks. The go-web.bat file not editable. That's a red flag. It could be false positive but nevertheless be careful of bad actors. The OG author's builds are editable and don't raise any red flags. All of the source code is readily available for you to look at and probe for any external connections which could be hidden. I did not find anything fishy in the OG's builds after scanning every .bat and .py file with my eyeballs.
Thank you again! This tutorial has more details than the last one you made. A lot more is explained and I really appreciate this tutorial that you have made.
Glad it was helpful!
Nice.
The future of fully-voiced video games, regardless budget is near.
Yes, just listened to my trained model blasting out Bob Dylan's like a rolling stone 🙂
Sweet! 😉
aweseome stuff. i will definitely try it out. my hardest challenge would be to restore the voice from my gradparents etc from old restored reel2reel tapes... i barerely get 5 minutes or so in medium quality. lots of work, but definitely worth it. thx for bringing this up and explaing this stuff in very detail. i absolutely love your content. i already made my grand.grand.mothers picture saying something, which lead into a spontaneous flood of tears ... :))
Yes, hearing a loved one again can be a very emotional experience
Crazy !
Thanks for the thorough tutorial! It seems remarkably straight forward to do this now, even the training!
Yup. It’s super simple!
Sick burn on the Frechies. Really great results with this!
😉
Lots of great info. Always amazing all the efforts that you put in helping others. Cheers!
I like the part when you sing. Hidden talents 😂👍
Glad it was helpful!
I realise that You do not have to sing for the training DATA, but I did and felt a right prat haha.
I live in a block of flats with neighbours above and below.
Thankfully I have a spare room that's pretty well loaded with boxes etc, so not too reverb(y)
Anyway, I got a nice clean recording of me singing about how I'm singing solely for the purpose of creating training DATA. And other mundane things like apologising to the neighbours, that it would only be for 10 minutes etc 🙂
I used my trusty Tascam H4n Pro, and wind shield
My PC is just churning away now, and I'm looking forward to experimenting with this.
Thanks Nerdy for reminding me again about this project.
I've been loving your videos and humor! Been watching every video. Thanks for covering this ai voice stuff!
Thank you - glad you’re having fun! 😉
Tortoise killer? Rodent explains details...
Is just another technology voice to voice and not text to voice
@@ratside9485 ahhh honestly haven't watched because I'm working, but so vits killer😄?
@@wakegaryit’s better than tortoise.
Unfortunately when you run the AMD/Intel version you encounter the following error: "NO GPU DETECTED: falling back to CPU - this may take a while". running on your CPU is extremely slow and will take days. Please gives us an update if you find a way to make this actually work for those of us with with AMD GPUs.
do you find something?
Nice work! So let me get this straight...I could clone my own voice reading a paragraph with specific tones, intonations, etc...and then I could clone a second voice and have that second voice read back my original recorded paragraph with the exact same specific tones, intonations, etc...?
Yup!
Absolutely marvelous!😀
Love your stuff - searching the internet for the best AI cover apps to use to share later. Thanks for your in-depth vid. You got a new sub and a like for your spectacular input.
Awesome, thank you! 😃
You absolutely rock
Shame the elements song didn’t think to put oxygen/nitrogen on opposite sides of a breath.
A good update, thank you, It would be nice if you could show what the different models do and how to use the tensorflow for windows users.
tensorboard is exactly the same on Windows 😉
@@NerdyRodentI see! Thank you for your reply. Could you please consider making a video about how to use it?
An error in cmd shows like that when i trying to convert
if data.dtype in [np.float64, np.float32, np.float16]:
AttributeError: 'NoneType' object has no attribute 'dtype'
why this happens
please help
Check your inputs for whatever you’re doing. Make sure they match the instructions provided in the video.
Thank you, you are an amazing person!
Thank you too!
neat! as always, this is beyond me... :D but I'll hand it over to my brother who might figure it out.
Lets see Weird Al doing serous songs! or whatever random things!
I have faith in your brother 😉
@@NerdyRodent hahaha dude you should have Ken Ham or someone sing some atheist song :D
@@DeconvertedMan like catapult pancakes?
@@NerdyRodent I mean thats not an atheist song... but... it would be pretty damn funny to have a bunch of fundy nitwits singing that XD
I tried just installing and unzipping and starting go-web, but it doesn't exactly work. The CMD starts, but it can't find the specified path and that's where it can't go further. I'm no coder and I don't understand python, but I would like to know what I'm missing since this is clearly not install, unzip, and have fun.
The unzip _is_ the install ;) Make sure you're using the right zip file for the hardware you have.
@@NerdyRodent Looks like I installed everything right, just not in the exact way you did. For instance, I extracted with winrar instead of 7zip the first time, and I downloaded the correct file, but not from the exact same place. Just followed the exact steps you did and it worked this time. Thanks for the response!
have the same problem- the huggingface page also dont have any zipfile in the directory anymore. i downloaded it from github, but "it can't find the specified path". is it dead?
Spleeter (from the people at Deezer) works pretty well to extract vocals from music.
The full UVR has demucs too, plus a bunch of other stuff!
when i copy and paste the training folder, nothing happens :(
thank you so much!!
wow !
Is installation in windows via command prompt similar to the linux method ? the usual pip pytorch etc etc
Also, is anaconda necessary to the process or can we make our own virtual environments.
Thanks a bunch mate
You can indeed use whatever python environment manager you’re used to!
thank you for the video, very good explanation. i have a question, my training seems to be doing about 4:30min per epoch on my 1080ti at 10-11 batch size (any higher doesnt seem to work) and like 50% GPU usage. is this what i should expect from this card? do you happen to know any good cloud solutions? thanks again
Sounds about right for a card of that age. Cloud options will depend a little of where you are in the world, but AWS, paperspace and others offer GPU systems
Hello!
First video I've seen from you, thank you very much for this excellent tutorial, helped me out a lot!
I've got a question though: Is the webinterface available from outside of my machine? Can other people access it? It seems as if I am hosting it, and the batchfile says "go-web", so I wonder if there is any security concern?
Cheers, looking forward to more wonderful content from you!
You can, but it takes a bit of setup for remote access
There is a docker version, I'm definitely going to try that, why go through the pain of installing pytorch etc when you can just git pull and docker compose up --build?
Freedom is installing the way you want to 👍
@@NerdyRodent I second that 100%! I just have too much going on on my PC and I'm always scared of these things interfering with each other, that's why I love containers
@@reezlaw for me, all those docker builds take up too much space, which is why I prefer anaconda managed environments. Options are great!
@@NerdyRodent that's a fair point
so how do i use/setup tensorboard to check if im overtraining and when and how should i stop the training of a voice cause i have no clue what im doing
Hi nice tutorial but I've a problem.
Completed the training process no .pth file appears in the weights director. What could be the problem?
Thanks
It could be that it's actually still training, or there was an error during training. Did you see it counting through each epoch as it trained?
Do training files have to be in english or can you throw any language at it? And when you extract voice from a music clip, can you also take any language?
So far, I’ve only tested Languages which exist
@@NerdyRodent So, just english?
@@foreverjune8 😉
Been using RVC for a couple weeks and the accents aren't as strong , voices kind of mimic the accents of the original singer even when set to max.. any solution?
Fab tutorial as always! Thanks 😁
Btw I tried following your previous RVC tutorial and I couldn’t get good sounding results like in your demo. Not really sure what I did wrong, maybe a poor dataset but the voice was all screechy and distorted and robotic.
Will try again!
Watch for over training with poor quality data sets
i was wondering if you would create a tutorial on the txt2speech repo of MRQ ai voice cloning, its toriose doesn't take much time to train and the output is decent and theirs a lot of settings to play around with.. though not knowledgable and how to get the best results, it would be nice if you would go into that in ur tutorial if it felt necessary, but its a cool workflow to go from MRQ txt2speech to RVC speech2speech to get enhanced results, it brings things much closer to 11labs quality... but not in convenience for now
Yes, tortoise / vall-e-x / bark -> RVC is pretty good
Man, this is tough. Could you break down the process a little bit finer for an average windows user such as me-self?
Which step are you stuck on? 😉
the audio player doesn't have any way to save the file....any ideas?
I just right-click and save audio as
:)...thanks i did find it eventually....just being a num nut@@NerdyRodent
Can we do voice cloning using this above method and previously i was trying with a google collab project and i got the some runtime error. Does this project includes any google collab and stuffs clarify this. Thanks in advance.
There was a colab, but this isn’t really what colab is for. If you’ve not got your own computer to use, there are numerous rental options such as paperspace, AWS, etc.
I have completed training and given one click training..but where can I find the .pth file
@@gokulkrish3839 weights directory for the pth files
I have checked weights folder but after giving oneclick training.
Donno where the .pth file falls
@@gokulkrish3839 should be in there once training has finished, or the steps along the way if you enabled them.
How did you access that page on the hugging? Just found the regular "pretrained folders" and not the file structure that you show in the video.
I just clicked the hugging face link on the GitHub page, which is the same one as in the video description?
@@NerdyRodent i.imgur.com/vWnkX8k.png Was this page i oddly couldn't find. Which looks to be models trained by others as you noted :)
Edit, found it out, just search for "rvc", and they all pop up xD. Sorry for wasting a bit of time, just takes a click or 2 before i get "basics" :P
Can you kindly tell me how can we replace the vocal or the instrumental section of a song with a piano voice ? I really appreciate your reply🙏❤️
Hey, Nerdy Roden! How is this differnt than Bark AI you did a tutorial a while ago? Is this better? Also can it be used for something else besides songs? Narration for example?
RVC is the best open source singing voice converter I know of at the moment…
Thank you for your tutorial Nerdy, Im wondering where can I find the .index file after training?
It’s in the logs directory
@@NerdyRodent hmm, can't find it. Got a couple of .pth files named D_2333 and G_2333 which are 800 MB and 400 MB but that's abou it.
Training is not done yet, so maybe I have to wait until it finishes.
Yup, that’s the place. The index might have failed, so train just the index while checking he output logs
got it! thanks!
I'm having a lot of issues to actually install the dependencies. The pip install returns a non pip related issue
Not had any issues myself, but the full error message should indicate what went wrong. Make sure to use the correct requirements file for your hardware.
and what is the module do you use in your cover songs
How do we get our own voices? I only have mp3 files
I can't figure out how to get this to run on my Mac M1. Is there a very basic step by step of how you run this? I have no knowledge of python or any of this environment
Download and run the shell script as shown. Done!
I am not having any success with the Visions of Chaos build of this on my PC though your video is the first I have seen that explains the GUI and confirms I am using that correctly at least so that's not the issue, no failures just ...nothing, it never completes the first step
Probably best just to use this version. It gets updated pretty much daily, so anything from VoC will just be outdated.
I need a nerdy brain though
@@madcatlady 😆 you can do it!
Yees, finally a new vid on this
just have a question. How high is your batch size, when you train? Is it something that if you set it too high, you get an imprecise module? If I have a dataset of one hour, what should my batch size be?
@Nerdy Rodent, I have extracted the vocals of a pretty heavy death metal song. The vocals are scuffed a bit but thats fine. My issue is that the instrucmentals have alot of static, i used HP3, should i just try the others?
HP2, 3 or 5, yes. As mentioned, there is always the full ultimate vocal remover, which has a number of other models too
@@NerdyRodent Oh yeah, that totally worked better. There was still some echo or reverb (not sure). Honestly what ive first created is really cursed because Mr Krabz doesnt like harsh vocals but i think it makes it even better for a first attempt... Anyway ill try to tweak tomorrow to see how i can make it better
Hello sir! Great video :) But I need help. Every time I try to clone the voice, an error occurs: "nonetype' object has no attribute 'dtype'". And I don't know why. I name the file well (including the .wav), but it always marks error. The weirdest is that some few files work well, but most of files don't work
“NoneType” means there is nothing there, hence it has no attributes. Make sure your inputs actually exist, and aren’t of the wrong type if they do.
Thanks! This is almost what I need. However I want to replace the original lyrics with my own and vocalized by f.i. Elvis. Any chance of this happening with this software or some other?
Yup. Just sing whatever lyrics you want and then convert that into any voice!
I wish it was that simple but I can't sing. Compared to me rappers and Tom Waits are opera leads...
Auto tune to the rescue? 😉
@@NerdyRodent Yep, Elevenlabs, Audacity, autotune and rvc combined might do it. I still wish there was a single software to map text/speech to a tune!
is there any example how to consume it directly or through api without gradio ?
thanks, but anacconda seems to nor download for me. are the other alternatives compatible do you reckon?
Only the two provided, I’m afraid! Check your internet connection if files aren’t downloading properly.
what's the 2nd? And my internet is fine. (btw fast replies wth thx)
@@NerdyRodent
Zip file first, anaconda and git second. Either method should download all files correctly. Other things that could cause files to not download would be things like faulty anti-virus products, faulty drive, etc.
thanks it worked. without anaconda.@@NerdyRodent
in the new version of RVC, which has just arrived you have to put your modules inside assets/whaights. This is a problem when I have now made zip files that I have repackaged, so log and wheights are in the same zip file, so no assets folder. Can you change it? Now it detects itself which index file belongs to the module that you have now loaded
I'm on an Intel Mac. Been working at this trying to get this going for a WHILE. Finally got it running, but now it's seeming like I need a GPU in order for it to work on my computer. In the script one of the last things it says is "move model to cpu". How do I do this?
I’ve not got a Mac, I’m afraid
I keep getting "Connection errored out" after a few seconds of doing anything in the web UI. I haven't closed the command prompt, and gave exceptions for the antivirus. What could it be?
I guess it could be a false positive problem with the antivirus?
Hi. I have a background in audio production & want to ask a existential question. Creatives minds are using these tools in positive ways but what about the negatives.
.
Do you think it's possible to clone a voice so the results will null when compared with the original source material. Will voice verification tools be
mandatory in the future?
Perhaps one day, only time will tell!
Do you think it's possible to use this on my Hackintosh with AMD RX580 or should I go Windows directly?
Linux is usually the best bet with AMD cards, but unfortunately I know nothing about which cards are too old. Maybe give the Windows one a go and see?
@@NerdyRodent Thanks for the fast reply! I'll try Windows =)
There's this song written by my favourite band (A) that was sold and sang by a different vocalist (B).
What I want to do is extract the vocals of (B), create a voice model of (A), and replace (B) with (A), so that the original writer actually sings this song.
Is this possible?
Yup
I'm currently having difficulty converting voice on a song since it keep giving me has no attribute "dtype" error and I don't know how to fix it
Make sure you've got the path entered properly
RVC got banned from Google Collab a few days ago ! Can we do a video about that ?
I knew there was a reason I don’t use colab… 😕
@@NerdyRodent Same for me , I never trusted it ! There was always a risk someone can make the colab stop working and there is nothing you can do about it. That is the reason why I quickly preferred to install the RVC GUI for MAC and it works so perfect for inferencing ! However I wasn't successful to get the training part to work correct. I am still searching some support how to do this but everyone seems to work on PC. Now that the colab training has gone I will need to searcher harder how to get the MAC training done for RVC GUI, but thanks for reply !
Wait does it let you sing good though ? Like hit notes
Thanks for the video i have a 4090 and noticed with a similar sized data set if i set the batch size to 40 it was about 5 seconds a epoch but if i set it to a multiple of 8 like 32 it was giving me epochs in 2-3 seconds. Also the readme mentions access to v3 but the ui only showed v1 and v2 as options?
Have you tried the real time voice changer they link by Okada?
Yup, the real-time option is fun 😉
you said something to the effect that you could train for more than 5 seconds on an epox, how do you do that?
Yes, it takes about five seconds per epoch for that specific data set on my hardware.
@@NerdyRodent and what about the asset can i chance my location of my weigts
what mac do you have and what graphic card do you have?
I’m using Ubuntu with an Nvidia GPU. Unfortunately, I have neither an AMD card, nor a Mac
when one click training finishes theres no new files in weight what do i do the index files are generated in logs tho
Go back and make sure all the steps completed without errors
There is no need to stick with just voice. This can clone and generate anything. Cat meow to saxophone? No problem!
Indeed!
man, so i used the google colab, rvc v2 thing for a while now, and it was great, but now it just disconnects when getting to training step, this does not have the same problem?
Do you know a way around this problem?
Not sure as I only use it locally 🫤
RVC got banned from google a few days ago !
it was updated today! testing now, will update after test.@@klaurcschwackerberg1880
-edit
seems to be fixed maybe? idk its different
@@IAMERROR64no they still banned
Does anyone know the command to start in the dark theme? "--dark theme" doesn't work.
Hello, I cannot train a voice, the site does not recognize the address where the file is located, I tried everything, changed folders, .mp3, .wav, .pth, nothing there do ! Especially since I already have the .pth file, can you help me please?
Drop me a dm on www.patreon.com/NerdyRodent and I’ll see what I can do!
Do I still need to make each audio file length less than 10 sec for the dataset like the so-vits one?
Nope 😀
Think i have to wait for the author patches it out. Using stock voices, and hit convert, and i get this
AttributeError: 'NoneType' object has no attribute 'dtype'
And error above saying " ffmpeg._run.Error: ffmpeg error (see stderr output for detail) " And Can't seem to find this stderr log.
Probably worth checking your audio data set as it’s saying it doesn’t exist (NoneType)
@@NerdyRodent Aye. Got it all working now, just took miniconda and made a venv instead :) And tested a 10 min audio sample, and 200 epochs was not enough, so doubling it saving every 20 epochs until i find the best ratio :P
win!
@@NerdyRodent Yep! :D By the way, do you know of a better text to voice than bark that can read pre trained voices in .pth format like the one in this video produces? As bark also uses seeds which is kind of annoying as i'm lacking a module to set a manual seed, and lacks "use last seed", and it's results is quite hit and miss.
How to use in amd card any suggestions I tried amd version but it uses my cpu...
No idea - I only have Linux + Nvidia because I like to have the most compatible, easy to use and supported AI setup :/
Hi, which rvc version should I download if I am on a M2 Mac ?
Personally I’d go with grabbing whatever the latest version as it gets updated quite a bit!
@@NerdyRodent Thank you. I got it to work and process. but the result output audio file is all in silence, has that ever happened to you ?
@@brucejsg not had any issues yet!
Thank you for this video! Your video was very helpful, but when I use the female voices, they don't sound at all like female voices; instead, they sound robotic-like masculine voices. How do I make them work properly?
It could be they are overtrained?
Oh, I think I figured out the issue, and the female voices now work fine. All I had to do was change the transposition value to around 10-12 instead of the default value of 0. Regardless, thank you for your quick response!@@NerdyRodent
Hey, trying to separate the vocals and instrumental using the GUI, but I'm getting an error every time despite following the same steps as you. I could paste the output information here, but it's quite lengthy. Think you could help out?
same problem did you find a solution?
make sure you have no spaces in folder and file names it works for me now
I'm really not understanding your guide for the portable version, I don't have the asset folder when I download the zip, and even in your video you don't have it either then you say to jump to the next section and you all the sudden have that folder, am I suppose to download that folder separately or what?
The older versions do not have an assets directory yet
@@NerdyRodent i appreciate the response, but are you saying the portable version isn't updated? or the older models don't have an asset directory? Am i suppose to do something different in order to use a v2 model? Sorry if this is basic knowledge, im just having a little difficult time understanding it
That’s correct. The portable versions are as old as the date you see on the right hand side on the hugging face website.
@@NerdyRodent Alright, that makes sense. Thanks again for taking the time to answer my question
I have an IMac and using windows with parallels desktop. AMD. Will it work?
I have neither a Mac nor an AMD card, so I’d say give it a go and see! Inference should be possible on MacOS
this is the daft punk vocoder model. How can i train this? altso with out v2
Output information
Model information: 200epoch
Sampling rate: 32k
Whether the model inputs pitch guidance: 1
Version: None
Better gpu for stable diffusion sdxl rtx3060ti or rtx 3060 12gb?
More VRAM = better AI experience
How do you see te tensorboard?
Assuming you did the normal install, run the tensorboard command as shown at 22:21
@@NerdyRodent Hmm yes i saw that but i'm cobfused what window that is, wehere it's from.
@@kenz2756 it’s the same terminal used throughout the video. Just run the tensorboard command in your RVC environment.
@@NerdyRodent Ah I see, you used a terminal to launch rvc in the web. I just use a batch file already available in the directory.
Yes French is the best lover language ;)
Yeee. I'm like, is he joking? It's one of the top languages. Enjoy took from various languages, too. I'm part French. Hehe. :3 22% in DNA. I remember the Dexter's Laboratory episode of French speaking back in the days when it came out. Hehe :3
it doesnt let me unzip the files...
Do you know which operating system you’re using, and have you ever unzipped a file before?
@@NerdyRodent I’m very new to this , I just bought my stationary setup. I tried right clicking in the 7zip but it didn’t show up - as for operating system , I’m not sure , I have windows 11 though .
Windows 11 (assuming you’re fully up to date) should be able to unzip 7zip files ok as of Build 23493. If you have an older build, then you will need to download the 7zip program as Microsoft Windows is too out of date to support modern files, such as 7zip. Go to www.7-zip.org/ to download a special program to make windows able to unzip files.
@@NerdyRodent Awesome , thank you for the help ! You’re amazing !
What is this xformers error? How can you fix it? Or does it have no effect? Thanks for your work
You can just ignore it. Triton is a Linux thing.
Ok thank you. Danke 🙏@@NerdyRodent
What exactly does rmvpe do if your just trying to get the best voice, is it better than harvest and crepe?
Yup, much faster and lower resource requirements too!
Thats great! Also I am not using any pitching correction I just do all that in audacity before converting, so if I do that, then is rmvpe still better quality than crepe and harvest? And thanks for the reply and awesome video tutorial, cheers!
it says error when i convert
Make sure to read the error and then take the appropriate steps to resolve it
@@NerdyRodent Already got it. thanks
Hi guys! Is it possible to install on MBP 2015 i7??
NO GPU DETECTED
What can i do?
Which GPU do you have?
@@NerdyRodent AMD RX 570
My first guess would be that card is too old
I think this source should also be on mobile devices, which I am on one, I haven’t heard of this source before, so I want to try it out.
Is there something like this combined with TTS?
Yup - audio webui
@@NerdyRodent thanks! Did you make a guide on this one?
the portable version did not work for me.
Windows 10, Nvidia 3080. error importing numpy
"Importing the numpy C-extensions failed. Original error was: DLL load failed while importing _multiarray_umath: The specified module could not be found."
I posted it on github, but in case somebody knows the solution for the portable version, please let me know...
You can just do the normal install in that case
I also question french being a legitimate language.
what are the requirements
No way to use AMD GPU on Win 10 ? I have an Asus RX 6700 xt but the RVC didnt detect it :|
AMD GPUs are supported best under Linux, so not sure about Microsoft Windows 🫤
Can it clone normal spoken voice as well?
Yes, as mentioned, you don’t have to sing at all
If only there was a mobile version of this 😢
Yeah, kinda needs a GPU at the moment
12:34 -> i started melting
i was hoping to listen you sing at the end of the video... lolz
Some day can you please make a dedicated video on how to install multiple AI tools and save Precious HDD/SDD space too.
At this point we have many AI tools, each requiring different packages but the same python version and for some AI tools a different version and combination of requirements. Is there a way to avoid multiple installations of the same GBs worth of packages and Just point to the existing under Linux?
Anaconda does a fairly good job for me, certainly when compared to docker or portable builds
Thanks 👍