Okay, back with me Arda, because many people have experienced errors in the previous version of RVC, here I will make a tutorial for you to solve the error. Now for this video, I will also provide a timestamp in the video description for those of you who want to jump straight to the video section. Specifically, you can check the timestamp below. For the RvC tutorial itself, I've made 3 previous tutorials. First, I made an RvC tutorial for the first version. Then for the second version. Then for
the third, I made a tutorial to make space on Hugging Face, like I did. I have and now I have to make another tutorial because the previous rvc tutorial can be said to be expired, why is it expired? that's because now google collab started banning google collab notebooks which made the google server expand, so at first google did give us free to use google collab but when the google server exploded I mean it exploded So Google did give us free access to borrow their computers through Google Cola
b But of course this has a limit because Google's own Google Server must have a limit, right ? use Google Collab Free to host a website or Gradio or Web UI and the like , where in the previous revision tutorial, we used Gradio, yes, the public link that you clicked at the end, that's banned by Google , so for those of you who used my previous Google Collab , after you guys running Gradio a few seconds after that you will definitely be cut off by Google . For the quick alternative, actually you c
an join Collab Pro right away . this video is all the way to the end so in this tutorial I will use keggle so keggle here is more or less like google collab they both give us to borrow a computer for free, but in terms of the features themselves I still prefer the features of google collab because we can control the connection between our Google Collab and our Google Drive account while on Kaggle this can't be done or maybe I haven't found a way or what but for the time being I still haven't fou
nd a way and this is enough to make it more complicated For the limit itself, That is, if you have used Google Collab , you can be hit by a daily limit, when you have been using Google Collab for maybe 3 or 4 hours, a notification will appear if you have hit a daily limit . per week you will be given access to access the keggle for 30 hours If you have exceeded 30 hours, you have to wait for it to be reset the next week Or you can also create a new keggle account But unfortunately here to create
a keggle account you need to verify your cellphone number So if you want to create a new account , please prepare a lot of cellphone numbers. And here, thanks also to Frieza for helping me to convert our notebook, which was originally from Google Colab, so that it can run on Keggle. First, before we go to Kaggle, You have to prepare the dataset first. For that dataset, you need to use this Google Collab Queue . So for this Google Collab Queue , it's not subject to a new policy from Google. Beca
use to create this dataset, we don't host websites, web UI, or Gradio here. So for this Collab it's safe to run, you can still run it for those of you who are new here, you can check my video here in this minute until several minutes later for the results it will be on your google drive for example here is my example there is this kafka folder in the dataset folder there is this kafka and here there will be several audio.wav files yes then you guys go to your description open my keggle link this
one is the rvc version 2 keggle you click on it then on the top right if mine says here it says edit but on yours it should say copy my edit or something like that yeah you click on that to copy my notebook then a display will open a notebook like this is on keggle, so it looks more or less the same as on google collab but in terms of its features actually I prefer google collab than this keggle Well, the first thing you have to do You go back to your Google Drive So, For example, I wanted to u
se Kafka earlier, right? So, you right-click, then you download it. Later, after you download it, you will get the .rar file, .zip file, and we will upload the zip file on keggle , so that's how you go to the section on the right, for example, for example, there is no right section, you can click on the lower right, there is a small arrow to open this data section, so click on the upload data icon for the name, it's free. If I'm here, I'll name my dataset as kafka voice, right ? then you upload
the zip file that you downloaded from Google Drive earlier, so the weakness is here, guys, so if you use Keggle, you need to download, then upload again, download upload download upload and it will definitely take up a bit of quota for my friends who use quota, because Google Collab belongs to Google so we can directly connect our Google Collab with our Google Drive account but here it doesn't seem to work. Now after you upload it, all you have to do is click the create button. But because I've
already created this dataset, I won't be creating it again. upload you click on the add data section here Then click on your dataset Now here you click on the dataset that you added earlier So I just clicked on the Kafka voice Then you just cross it now here when it's done it will display the input file here there is a kafka voice, in it there is a kafka folder and if I look inside this kafka folder there is audio in the wav format which we have prepared on google direct earlier now we run it fo
r the keggle so first you have to make sure the accelerator is here or you are using GPU which is GPU P100, yes , I used the T4 GPU but it got an error, so my suggestion is to use this P100 and for the quota limit, you can see below, so I used it for almost 6 hours to type on this notebook so it can run and the weekly limit is is 30 hours, yes, and here there is a reset date, which is a reset on September 2 , then what you need to do next is for the mode, because here we are doing training, let
it be training , and then for the model name, it's up to you, but you have to remember it, this is You have to remember the model name later, at a later stage, you have to know or remember the model name that you are using, so here for the model name, I will use Kafka because I use Kafka's voice. After that, all you have to do is press the run all button above, click run all , here for the process, maybe 1.5 times or 2 times slower than google collab, yeah, because I think the CPU used in Keggle
, maybe yes, the specifications are lower than those in Google Colab , so for the process of downloading libraries, installing libraries, and so on it's indeed here from my point of view it's slower okay so I'm here to start yes earlier run all from around 22.02 at 22.02 we'll see what time we can finish it okay so we need more or less 5 minutes to run the kegel until this appears in the webui run section here the public url appears here if you guys appear some warnings especially this 6p warnin
g yes , you can leave it all this time, the public url is still out now, all we have to do is click on the public url that was given, yes, we click on this Gradio , so this is Gradio, this is what Google actually banned, so Google wants you to use the collab, you can use it for coding on this page, don't you host Gradio or something like this and you leave your notebook instead, you guys are playing around in Gradio like this, don't do that, Google actually banned it. So, we've come back here to
the Gradio. Then we go to the train section. Because we want to train, right ? For the name here, make sure you have the same as the model name earlier . 48k is better, so the version here I also recommend for version 2 because it's even better here for the training path of the sound that will be used or the dataset that will be used we keggle again and then this kafka folder that we entered earlier we Click copy directory path, we will return to the Gradio, then we will paste it here. We can c
lick on the data process while we are waiting for it. For this part, we don't need to change it, just leave the RMVPA GPU here for the total training , it's free, it's up to you , but because I'm here just for test so that I don't stay here too long I will just use 5 epochs to save the frequency so save this frequency, for example I use 50 epoch so, this save frequency function later every 5 epoch it will save our model so 5 save epoch, 10 save epoch, epoch 15 save, epoch 20 save that's actually
good for your backups in terms of backups it's good, but what's not good is in terms of storage so your storage will eat up a lot for myself, I usually set this epoch how much I want because now for my test it's 5 and save frequency I usually don't touch it but for this section save only the latest I usually click yes so only the last file will be saved, the last file so it won't take up a lot of storage but if you want to look for it it's safe to save You can actually maximize this frequency s
o that it doesn't take up a lot of storage later. Then to catch all this training , it's the same as the previous tutorial. it's better to click no because if you choose yes I'm afraid it will actually go the other way it will burden the GPU and it will actually become an error so because I'm under 10 minutes I click yes then we go back to the top, in the data process we have already clicked if we scroll down , if there is an end process, that means it's finished , we can also check our kegel se
ction, if we scroll down our notebook, here is also the end process, so the data processing is finished, then we click on the extraction feature. Now here is the extraction feature. It's also finished, so the marker is that if you scroll down in the Output Information section, it says All Features Done and if you go to the notebook, you scroll down too here All Features Done Then continue to the bottom, we first train the index feature first Now here, if output information like this appears, it
means that the index has been created, it's done then now we click on the trend , this model is a bit longer, but if I'm only 5 epochs, it's a while. Well, here it is again, the minuses and I still haven't found a solution, so the minuses here are what I found on Google Colab , I mean, what I found On Kaggle , when we train for the epoch, you can't see it , so the output is here, we can't see where the process has gone now , how many epochs have you reached, but there will be a marker later, I c
an show you a marker when your training is complete it's finished but for the training process where has it come to while I still haven't found a way how to see the training process now here usually when your training process is a marker yes it's a marker that your training process has started here usually at the end it says calls into the C++ blah blah blah blah blah for the GPU here the usage of the GPU has gone up, it 's already red here so it can be seen that the GPU is being used for traini
ng, right ? I found an error or what or your training didn't work like that, right? Then, the graduate section should also be here going around and around while we're training. But I don't know, because I only have 5 epochs. There's a possibility that your epochs here are large . an error appears And actually it can't be used as a marker either It can't be relied on as a marker so when your training is over, the marker in the grade appears Chinese writing like this if we translate it after the t
raining is over, you can check the console training log or the training log under the experiment folder , that is, if we translate this, R is finished, if it's like this, it's a sign that in your own Lognu, there's already this writing , this is a sign that your training is over, the GPU doesn't go up anymore to make sure it's over your training has finished or your training has an error like that, right, we don't know, to make sure it's actually finished, there's no error, you go to the output
section here, you click like this to expand the folders, then you go to the asset section here, you go to the asset section here. weigh now if your training is successful here in the weigh section there should be kafka.pth then if you go to the log section then the kafka section is also here, it should already have this, edit blah blah blah blah blah. there is also this totalv.npy now if there are three of them it means your training has been successful now we can stop the webui we stop it then
it will immediately run the code below to make a zip file containing the three pth earlier , npj, with the index, here we just have to wait for the process, now, if the successful zip appears here , kafka.zip will appear here, in your output section , here the size for mine is 85mb, so here's the drawback, yes, we have to download it again, so this zip what's finished is that you need to download it , so be enthusiastic for you quota users, now to download it here, it's not visible, so it's like
the download is running in the background and we don't see it, the download is already in the process, how come it suddenly shows up, the download results just like that, so yeah My advice is to be careful when you press this download, don't download it, maybe double, you can download it 2-3 times later, I'm afraid, now it's already downloaded, I'm kafka.zip, I want to rename it first, so there's no number 1 . training, next is The Inference Now for the Inference, I don't really think about it
now, I don't really recommend the Inference here, and the method is a bit complicated, because you can't directly connect to Google Drive like in Google Collab yesterday, so I have another alternative. you guys and at the same time this is also the result from what I tried, the result is much better than the usual inference that you guys use here, okay, for the first method, we go to hugging face, hugging face, we click at the very top , you guys don't have an account, please create an account,
I don't need to create one tutorial on how to create an account after your account is created, you can go to the top right on your profile, then click on this new model so in this hugging face we will save the model from our training earlier so the name here is free I'll just name it Kafka and for this make sure public, then create the model, then you go to the file and version section, here you click on add file, then click on the uploaded file, here you just drag and drop the zip that you down
loaded earlier, the training results are here, I drag the kafka.zip, we click this first, commit change to main, now it's just going to start the upload process, okay, this is for progress, you see, so what is this function for? This is so that you only upload the model once, so okay, you upload it once, let it be on the hugging face , because if you don't upload it on the hugging face, every time you make an inference, you need to upload it again , upload it again, it will eat up a lot of quota
and also time, ok, now you go to the link in the description again, you open my kegel which is AI Covergen, right, and this is for the notebook, it's not mine , it's made by a YouTuber, it's called socialinepweb, right, and it's the same as before, you guys click here, you should copy my edit on you, OK, you click on that, so this is the original code that has this code, this socialinepweb, so we appreciate him who has provided us with this code, what I did was just put the code in this kaggle
so that you can use the accelerator here, this time I'm using this T4 GPU and oh yeah, make sure this is your quota, yes, your quota doesn't run out for the previous notebook, because we've finished here , we keep turning it off for AI Cover Gain , we just run all , so While waiting for this to all run, I want to explain to you guys a little why the inference results here can I say it's better like that, right here because I'm a lazy YouTuber I just use paint so why is it better there because la
ter in the gradient we only need to enter Just the YouTube link, so first, we'll enter the YouTube link, right, the YouTube link that contains the song that we want to cover, then the song that we want to cover, the program that was made by socialist inepwip, the program will divide the music into 2 parts, so he divides it into 2 so it's vocal and also the instruments, so after he divided it into 2, it didn't come to this point that he divided these vocals again into playing vocals and also back
ing vocals , so he divided them into two, so playing is the main voice of the main singer's voice , this backing vocal is the voices of people's voices also, but this isn't the voice of the main singer, like the voices of the instruments that don't belong to the voice of the main singer will be divided like this, so there's playing , there's back vocals. hit, got the name of the reverb so what's in this reverb? so there are some songs that have echo effects from the singer, right? Usually, songs
with this echo effect are hard to detect by AI and that makes the song cover look ugly , this developer, the social line web this is how it's applied to the reverb here so the song that's been split up so the main vocal needs to be reverbed, so the reverb in the song is removed, right after this, after reverb, the reverb is removed , then the next result is this, that's an inference with the model We've trained , so after inference, we can add this here, yes, add reverb, so our song cover resul
ts will be even better, after reverb, we'll put it together again, we'll meet this, we'll meet this, we'll put it together again with him and it's a song it's new , I'm lazy to make good ones, we just use pain , so the system is like this, the flow is that's why the inference results that we'll make later will be better , although for some cases there might be something missing , okay, because it's also called AI yes, it's still not 100% perfect here for covergen AI notebooks, we've reached the
bottom here, the public url appears, we click here, it looks like this, first we go to the download section, the download or upload model is actually the same, so if you want to upload a model, please use it The Kafka zip was earlier, you can enter it here so the minus is if you upload directly using this upload feature, so it's like every time you want to reference a song, want to make a song cover, upload it again, upload it again, upload it again, it will use up your quota, that's your quota
and for your time, so the easy way is with the hugging face earlier so we go to the download section of this model, then we go to the hugging face that we have created, this kafka.zip, right-click right- click here, right-click on the download icon , then copy the address link, after that we come back here, download the model link, we paste it here , then the model name is Kafka , then all we have to do is click download . access fully download then to check if it's already there or not you can
click generate you click refresh this model now here you click the drop down here kafka will appear if you've succeeded now but because kafka is just me trending 5 epochs the results are not good yes, it can't be a bit questionable, the results are so here I'm going to use the one provided below. There's an example. I just click on the funny one. I click on the funny one. It will go straight to this model link , then I download the funny one. Next, we need to prepare the song we're going to cove
r, so we're looking for the link, we're going to YouTube and this time I'm going to use the song I missed last night, okay? Here we copy the link address, then we go back to the grade . If you want to upload the file, you can click upload this file but because it's easy, we'll just go straight to the YouTube link, so here are some settings, there are a few settings on the right, so to change the pitch, there are 2 options because this is a girl's song and my model is also a girl, so this is refr
esh the model first , okay ? _ deeper if you go to the right, the sound becomes shriller , and the difference is for the one above and below , if this one is above, it only changes the pitch of the vocal sound, while if it's below, it will change the pitch of the vocals and also the instrument, but be careful when you play it the one below Usually the sound of the instrument will be strange if you give too much or change the pitch The sound of the instrument will become strange So this works for
If you understand the world of music, the world of audio So there is a key, key from a piece of music So it's like the key for a deep male voice is different from the key for a girl with a high voice I don't really understand it either but you can try playing this, the point is if you go left it gets deeper to the right it gets shriller but because it's here From the model it's a girl, the voice is a girl, I mean the song is a girl , so I'll just leave it zero and continue here for the voice co
nversion option for this index rate, like I explained in my previous video tutorial, so here's how much we want to use file.index, file.index, which we trained earlier what is file.index for? file.index's function is to store the accent of the model , so the higher the accent, the more visible the voice accent, but sometimes this index actually makes noise like that. your cover so sometimes this index is good given sometimes this index is good given 0 so you can just try it with your respective
models so each model is definitely different and I rarely tamper with this filter radius, I also don't understand this RMS Also, I rarely tamper with it, this protect rate is also rare, but there's also a description here, you can read it and you can try it, who knows, you might find one that's just right for you, right here for the pitch detection algorithm, I usually use rmvpe or you can use purcreep, also you can both be equally good because both of these are the ones that are being used by a
lot of people now here is what I said earlier, there is audio mixing here so we can adjust the volume for playing vocals, backup vocals and also the music, for example, after you make an inference, it turns out that the vocal sound is too high , you can lower it or if you want to lower the vocal backup sound, you can increase it, you can, and so on , so here you go, this is the control for the Reverb, you can also adjust it for this Room Size The assumption is that if you sing in an empty room,
it will usually make an echo , and this adjusts the size of the room, the bigger the room, usually the more reverberant, so 0 for no echo, right, this is the biggest room where the echo is steady here, for the initial experiment, let's leave it at 0. 15 yes, like the initial setting we didn't have to tweak anything so earlier I didn't have to tweak anything, we try to inference, to generate later we will, the second we will try to tweak it so we click generate now here the process will take a l
ittle longer compared to your usual inferences because, like I said earlier, there are a lot of processes here, so first, if you see it, it will download songs from YouTube, from the YouTube link that we have entered . Well, then it will separate between vocals and instruments. Well, then he will separate between playing vocals and backup vocals. So, like I said earlier. So, the main vocals with the backup vocals, which are usually behind the instruments, will be reverbed at the back of the song
, right? removes the echo sounds in the vocals, when playing the vocals, then he will convert the voice using rvc, now he just does the inference, then he will combine, yes, put it all together so it's like what I wrote here, so it's an inference, give it a mix, then put it together again here for the results maybe I'll just rotate it a little bit, I'm afraid it will get copyrighted later so this is the result by default, I don't change the mixing I love you, you love me too, but honey... So tha
t 's the default result, here we try to increase the room size . where are the numbers from? Actually, this number was made by me trying it myself So I'm not sure that every song is always good at this number You can try it yourself Then we generate again And this is the result I love you too, I love you too but I love you too but I I love you It's a shame too . That's it, right? There's an echo like that, right? After mixing, you can play around with it yourself for the mixing. More or less, th
at's for the tutorial. This time, if you have something you want to ask, you can join my Discord server. If, for example, I can't answer quickly, maybe you can ask questions to other viewers who have joined the Discord server and for those of you who want to support this channel, don't forget to like, share and subscribe to this channel and if you wish, you can donate to my link . to treat or my saweri is in the link in the description below thank you
Comments