품질과 속도를 위한 AI 음성 클라이언트 설정 최적화

AI 음성 클라이언트를 최적화하려면 하드웨어를 조정하고, 적합한 모델을 선택하고, 속도와 품질에 대한 설정을 미세 조정해야 합니다.

### 속도와 품질을 위한 AI 음성 클라이언트 설정 최적화
* ⚙ AI 음성 클라이언트의 성능은 하드웨어의 영향을 받으며, **구형 또는 낮은 사양의 시스템은 실행 속도가 느리고 끊김 현상**이 발생할 수 있습니다.
* 📥 AI 보이스 체인저는 이제 **Hugging Face 저장소에서 다운로드**할 수 있으며, **GPU, AMD 및 Mac용 버전**이 제공됩니다.
* 🎙 주요 설정에는 **게인(마이크 볼륨), 튜닝 팩터(피치 조정), 인덱스(훈련된 모델의 경우)** 등이 있습니다.
* 🔊 Onyx로 내보내면 **모델을 더 효율적이고 빠르게 만들 수 있지만 인덱스 파일을 사용할 수 없습니다.**
* ⚖ 청크 크기와 추가 매개변수를 조정하는 것은 속도와 품질 간의 균형을 맞추는 데 중요하며, **값이 클수록 품질은 향상**되지만 성능은 느려집니다.
* 🔧 클라이언트 모드와 서버 모드 선택은 **속도와 품질에 영향**을 미치며, **서버 모드는 속도가 빠르지만 품질이 떨어질** 가능성이 있습니다.
* 🎚 크로스토크 및 RVC 품질과 같은 **고급 설정**으로 음성 출력을 더욱 세밀하게 조정할 수 있습니다.
* 🖥 시스템 요구 사양은 그리 높지 않으며, **GTX 1660 또는 1060과 같은 저사양 게이밍 PC로도 AI 음성 클라이언트를 실행하기에 충분**합니다.

today is all about the AI voice client and its settings. I'm going to be going over how you can get it running faster, how you can get it running for quality, and how you can get a little bit of both. So let's go ahead and jump into the video. And actually, before we do jump into it, I do just want to touch up a little bit on the hardware for this. So everyone's going to have a little bit of different hardware, and there are some trade-offs. The older your hardware, the slower this is going to run. So you're going to have to run things in a slower fashion. The newer hardware or the higher spec your hardware is, the faster you'll be able to run it without it being choppy. So unfortunately, there's no way getting around that. There are some optimizations that we can do to try to help out these systems that are a little bit older. So just briefly wanted to touch up on that before we jump into the nitty-gritty of it. So there is one thing

I want to say and that is there is now a new link to download the AI voice changer. So apparently I crashed the Google Drive one because if you go to the Google drive, it says the quota has been exceeded. So what you have to do do is head on over to hugging face. And so we're gonna head on over to the hugging face repository, and we're gonna notice that there are a bunch of files here. So what works for my system is this bottom one right here, which says Onyx GPU Cuda. If you have AMD, you can try out this Onyx direct ml. However, the author does say that AMD is experimental, so it may not work with everything. And then if you have mac, you're going to have to select this top one right here, and you can see that on the right side of the window. So those are going to be the newest ones, and this bottom one is what I have here. Once you have that, it's going to be exactly the same installation as the previous video, and you can carry on with

the rest of this one inside of the real-time voice changer client. When you boot it up, you're going to be met with this screen here. Some people have been having issues with selecting a VC, a voice changer, on my other systems. I was able to just click RVC and get it running, so I'm not too sure what has been happening with some other systems out there. But once you select that, you should have four models here originally, but I have a bunch of other voice models I've loaded into here. So let's just go ahead and select one of these voice models. So initial settings gain, this is going to be how loud your microphone is. This input right here and this output is going to be how loud it outputs of the speakers. You can simply just leave these at one and one. However, if you do realize that your output is coming out too quiet, you can increase the output gain or increase the input gain. Either one works. In my experience, this tune Factor right here is going to be the pitch of the voice. So at zero, there is no tune.

adjustment. If you're going from male to female, you're going to want to do positive 12. If you're going from female to male, you're going to want to do negative 12. If you're going female to female, zero. And if you're going male to male, also zero. Let's go ahead and take a listen to this real quick. Okay, this is me using the voice at zero. This is me using the voice at 12. And this is me using at negative 12. And so there's one important thing that you might have heard in that last snippet of audio, and that is kind of how it sounded robotic and distorted. So that means if you're hearing that, either it's outside of the pitch of the model that's been trained, or the model wasn't trained enough on that pitch. So I went super low. Boton's model wasn't trained on very low pitch, therefore it sounded very robotic. So that is one way you can determine whether or not the model has been trained enough for that. The next thing

is going to be index so index doesn't really matter if you did not select an index. What do I mean by that? So if you go into this edit tab up here inside of the voice models, you can load code in an index. So I did actually show you how you can get the index in there in the previous video, but that is incorrect on how to do that there. So let's go ahead and do the correct way to select an index. You go into select, assuming that you train the model. This is only if you train the model and have RVC installed on your computer or from Google Colab. Once you're inside of the main folder, you're going to go into logs, and then you're going to go into the name of the voice that you trained, and then you're going to add this index file right here. It's going to be the one that's huge, and it's going to be added underscore ivf6106, and you're going to click open. Now that you have your model selected, you've got an index. You can now click

upload and it's going to upload the voice model. And the way that you can tell a voice model has been loaded with an index is in the command line window. You can see loading index here, and then if I click another voice that does not have an index, you will see that it says index is none and it's not used. Let's go ahead and hear a difference between the two. So here I am speaking with the index at zero, so it's not using any of the index file, and it sounds still pretty good in my opinion. And then here we have it at one. Honestly, both are good. You can actually put it in the middle at 0.5, but I just leave mine off because it actually makes things faster. And that's an important part right there. It actually makes things faster. So if you have your index at one, it's going to be using that gigantic file to reference things. If you have it at zero, it's going to be much faster as it's not using that index file to kind of judge how the

voice is going to sound so zero sounds more than sufficient. And all of the audio you've heard on my channel is at index zero. And there's one more part that I want to go over, and that is this export to Onyx functions. So what does export to Onyx do? Well, it actually makes your models more efficient and faster. Let's go ahead and click generating Onyx, and this window is going to pop up where it says generating Onyx file. Please wait a minute. It might be more than a minute. It depends on what your GPU looks like. So just give it a minute, give it some time to process, and then a new file exploit window is going to pop up. So here we go, a window popped up, and now we can go ahead and save our model. So let's just go ahead and save this here, and then once we save it, we can go over into our edit. Here we can go to select model, and then we can upload it. So now it's going to be using an Onyx model and with.

an onyx model. You can't use the index so now we have the voice in Onyx and the model still sounds good. Onyx is supposed to be faster than Pytorch files so that is this session right here. So let's head down perhaps the most important part which is chunk and extra. This is where you're going to do all of the fine-tuning for how fast to go and the quality and adjusting parameters based on what your hardware can handle. So to illustrate my point, I think it's easier to show you so that you can kind of get a feel for on how you're going to adjust it based on your system. What I'm going to do is start at 320. So this is kind of a middle ground. This comes with an 853.3 millisecond delay which is roughly about a second. If you take a look in this left corner right here, you can see the response time which is 9.73. If I go ahead and start it, you can see it adjusting here, okay. So here I am.

talking at 320 chunk size, and an extra of 496. So let's go ahead and decrease chalk to be faster. So here I am at 40, and you can see how it sounds kind of weird, is a little bit up and down. So this is sending smaller chunks to hopefully try to get the audio back to you fast too. If I hold a single syllable, you can sphere you can kind of hear the audio chunks. So you hear how the audio kind of sounds like a wave, that's because of the Chucks. Let's go ahead and increase it, but now we'll add subscribe if you're starting to hear some stuttering in the audio. If you're starting to hear some weird audio, your Hardware might have to go a little bit slower. So um, you go to 448, hear how that sounds. If it still sounds weird, go to 512, and keep doing that until you reach the largest chunk size to where the audio sounds smoothly. Now, of course, you probably, you know, using it in 2048 is going to be

rough because that's like a six-second delay. However, most hardware should be able to get away with it at around 512 and 448. So if you have relatively fast hardware, your hardware can handle actually sending more previous chunks. So what we're going to do now is we're going to increase extra. So we're going to bring this all the way to the max and hear the difference in sound. So here we are, a chunk size of 16 and sending an extra 130 bytes, 130k bytes. And I'm gonna bring it back down to 290s fixed and victims decrease in quality to about 32. Okay, and it's still decent. You can kind of hear it a little bit better. But if your hardware can handle it, run this at 131k and it's going to give you. However, these are the two parameters you want to adjust to fit your hardware. The chunk size is what you're going to adjust to get the audio sounding a little bit more smooth. And of course, if you go 4096, you're it's going to be less.

load on your hardware so it's all about checking which ones work best for your configuration. I don't think there is a one-size-fits-all here, just some general guidelines for me. 96 or 80 sounds just good enough with not that much delay for the videos. Of course, I do a little bit of post-production magic, but in reality, if I wanted to run it fast, I would run it at 24 or 16 or even 32 sounds pretty good. So, 32 isn't that bad compared to 96 on my system. So, hopefully I didn't make this section too confusing. You got to play around with a lot of that stuff there, and if you don't have a GPU, you run this on CPU. When I try to run it on CPU though, which might be reminiscent of something that I've heard people say that is occurring on their systems. So, let's go ahead and start it, and it's going to produce something that kind of echoes and kind of sounds weird. So, as you can hear, it's repeating. So, that is

not good right there. So what I'm going to want to do is increase the chunk size to let's say 320. And then we're gonna go ahead and restart the voice changer, and it sounds much better, all right? So the next part is going to be this left side right here. This one is um, not too important. You can either choose Harvest or crepe. I actually don't really notice the difference between Harvest or crepe. However, some people say crepe is better, but the author actually says Harvest is highly accurate. So the author says Harvest is the best in this case, but according to RVC, people say crepe is better. So just pick whichever you think sounds better. This is going to try to reduce Echoes. This is going to be like noise suppression right here. It doesn't work if you're on server mode. So, and the silence threshold basically adjusts how loud you need to speak before it starts converting audio. So if you're noticing your mic is too sensitive, I believe you can just

increase it all the way to 0.1, but you know, I just leave this down, all right. So the other part that is super important is going to be this client and server. So there are two options here: client is slower than server. However, I have noticed sometimes that client can sound a little bit better than server. The difference between this is server goes directly to this command line right here, converts it, and sends it back to your headphones. And so the setup is the same for input and output as you would for client or server. It just depends on what you want to use here. So client, you can now use the Echo and the noise suppression options, whereas server, you just send audio directly to the server, and you get it back as quick as possible. I like to run it in server as it's faster, and I'm able to get really good results still in server, but and there are two more buttons that we're going to go through in this window. That's going to be advanced settings. So here's this.

crosstalk button. I believe what this does is it changes how much of the audio is being overlapped, and this makes a pretty big difference when you're trying to decrease the chunk size. So to demonstrate this, let's just go ahead and try to run it at 16. And we're going to go ahead and do 1024 of a Crossfade. And we're just going to say a single syllable so that you can hear the so that you can hear kind of how the audio is being chunked out. So there's that with 1024. And then it's got a 4096. Ah, as you can hear in that audio, the 1024, the chunks sound a little bit harsher. You can actually hear the chunk. Whereas when you put it to 4096, it kind of smoothens out that choppiness in the audio. And then so for me, what I figured, what I found is Tran kit at 300 actually gives me a little bit better of that.

smoothening. I'm not actually too sure what Tranquet does, but I just do that at 300. And for Silence Front, I actually just leave it on. So, and then for Protect, I'm actually not too sure. So, I haven't played around with this one, but to me, it didn't notice any difference when playing around with this. Now, the last thing is going to be RVC quality. I recommend you leave it at low compared to high. If you put it at high, it just generates much slower at not that much of a better quality. So, I just leave it at low. I don't really notice too much of a difference. Um, you do get a little bit better quality at high, but I don't think it's worth it for the for how much it makes the system go slower. So, just leave that at low. And there is one more button, which is this Merge Lab right here. You can basically just merge voices here. I won't be going over in today's video. I have that planned for another one. So.

stay tuned for that where we will you know test out and merge out different voices so hopefully that answered some of your questions on how you can get this working on your system. It's going to be a lot of tinkering around with that chunk size and then the extra amount in order for it to get running smoothly on your system, and as I said earlier, this does require hardware to run. However, I believe you can run this on a budget gaming PC, um, you know, 1660, 1060, and from what I've been hearing, those are just fine. So, and so hopefully you found this video useful. If you have any questions, please leave them down below in the comments. I have been getting way more comments than normal, so it's taking me forever to respond to everybody, but I am trying my best to try to get back to you guys. So hope to see you guys later, and thank you for watching today's video.

Optimizing Settings in AI Voice Changer Client

I graduated with a B.S. in Mechatronic Engineering and here on this channel, I like to share my experience for learning and hope others can benefit from my mistakes and/or knowledge!  I share things I'm passionate about whether that be technology and AI stuff or anime and Japanese :)!  Thanks for visiting my channel!


Jarods Journey

AI 음성 클라이언트 설정 최적화

AI 음성 클라이언트의 성능은 사용자의 하드웨어에 따라 영향을 받으며, 구형 하드웨어에서는 프로그램 실행 속도가 느려집니다.

AI 음성 체인저의 다운로드 링크가 Hugging Face 리포지토리에 업데이트되어 Onyx GPU Cuda, AMD 및 Mac용 버전이 제공됩니다.

클라이언트의 초기 설정에서 마이크와 스피커 출력 게인, 음성의 높낮이를 조정할 수 있습니다.

음성 모델 조정하기: 피치 및 인덱스 선택

남성에서 여성으로 전환할 때는 양수 12 피치 조정을 권장하고, 그 반대의 경우에는 음수 12 피치 조정을 권장합니다.

오디오의 로봇적이고 왜곡된 소리는 모델이 해당 피치에 대해 충분히 훈련되지 않았음을 나타냅니다.

훈련된 음성 모델을 사용할 때는 인덱스 파일을 선택하는 것이 중요하며, 인덱스를 0, 1 또는 0.

5로 선택하거나 빠른 프로세스를 위해 인덱스를 해제할 수 있는 옵션이 있습니다.

인덱스가 있는 음성 모델을 업로드하면 프로세스 속도에 상당한 영향을 미칠 수 있습니다.

Onyx Export를 통한 모델 효율성 최적화

Onyx로 내보내면 모델을 더 효율적이고 빠르게 만들 수 있으며, GPU에 따라 성능을 향상시킬 수 있습니다.

Onyx 모델을 사용하면 오디오 출력의 속도와 품질에 영향을 미치는 청크 크기와 기타 등의 파라미터를 미세 조정할 수 있습니다.

시스템 성능에 따라 이러한 매개변수를 조정하면 오디오를 최적화하여 원활하게 재생할 수 있습니다.

원활한 성능을 위한 오디오 설정 최적화

하드웨어 성능에 따라 청크 크기와 추가 바이트를 조정하면 오디오 품질을 개선하고 지연을 줄일 수 있습니다.

청크 크기와 같은 다양한 설정으로 실험하고 GPU 또는 CPU에서 프로그램을 실행하면 최적의 구성을 찾는 데 도움이 될 수 있습니다.

왼쪽의 하베스트와 크레이프 중 하나를 선택하면 오디오의 정확도에 영향을 줄 수 있으며, 무음 임계값을 조정하면 마이크 감도를 관리하는 데 도움이 될 수 있습니다.

오디오 처리를 위한 클라이언트 및 서버 옵션 비교

클라이언트는 서버보다 느리지만 때때로 더 나은 음질을 제공합니다.

클라이언트의 에코 및 노이즈 억제 옵션의 사용 목적에 따라 둘 중 하나를 선택할 수 있으며, 서버는 더 빠른 오디오 처리를 제공합니다.

크로스톡과 청크 크기를 조정하면 오디오 품질에 큰 영향을 미칠 수 있으며, 청크 크기가 클수록 더 부드러운 사운드를 제공합니다.

오디오 처리 시스템 최적화를 위한 팁

최적의 성능을 위해 Tranquet을 300으로 조정하고, 최상의 결과를 얻으려면 사일런스 프론트를 켜두세요.

눈에 띄는 차이가 없을 수 있으므로 보호는 그대로 두는 것이 좋습니다.

품질 저하 없이 더 빠르게 생성하려면 RVC 품질을 낮음으로 설정합니다.

원활한 시스템 실행을 위해서는 청크 크기와 추가 용량을 조정하는 것이 중요하며, 이 소프트웨어는 1660 또는 1060과 같은 저사양 게이밍 PC에서 실행할 수 있습니다.