Optimizing RVC

[violet coral]

1. I downloaded following this link (https://rentry.co/forkvoicechangerguide) and followed it without an big issues.
2. The model stutters quite a bit and sounds robotic overall.
3. I have an NVIDIA GeForce RTX-4070
4. Perf seems fine; it's always been neon green
5. I've just been messing around with the settings but it hasn't made it any smoother nor any better. Oftentimes, the sound comes out distorted and sounds like a cat dying.
6. I've tested different models in this discord to check and see if it was a problem with the model itself, but the problem persists through every model.

So essentially, it's been stuttering a lot and I've been trying to use it mainly for discord. I don't see any huge issues for the settings after reading around, but it still doesn't seem smooth or satisfactory. Does anyone have any advice for this?

[grim geyser]

f0: rmvpe (onnx is mostly just for amd)

maybe you meant it's cutting off rather than stuttering? since the perf is green

are you using it for any games? is the perf not green during the game?

[violet coral]

I have an AMD Ryzen 9 7845HX CPU, but that shouldn't affect the settings

So I should run rmvpe?

[grim geyser]

yeah it doesn't matter bc you're using your nvidia gpu, you should use rmvpe for nvidia

[violet coral]

Alright, let me quickly test that out

Is there any problems with the Chunk and Extra settings that are immediately visible?

[grim geyser]

also, you can optionally set force fp32 mode on in advanced settings for better quality and stability at the cost of some delay

[violet coral]

Let me try that

[grim geyser]

well chunk is fine, it just needs to be a bit higher than the perf value which is, at max you could slightly lower it (still being higher than perf, since your perf is 30 and the chunk is 213 rn) to have a bit less delay, but it's not really needed

about extra, it should run fine, but if you're having cutoff issues, it's best to use 2.0 then

[violet coral]

Is there a way to type out a number instead of dragging the slider?

[grim geyser]

unfortunately nope

[violet coral]

😭

[grim geyser]

anyways you don't need to be precise for the chunk

[violet coral]

Alright, sounds good

[grim geyser]

soo, are you still having issues ?

[violet coral]

It's still cutting it off slightly

At the beginning of speaking and at the end of speaking

[grim geyser]

maybe set extra to 1.7, and be sure to not speak lower at the end of your phrases (you might do that since you hear yourself 'echoing' since you have monitor to headphones)

[violet coral]

I tried changing models and the voice sounds much more distorted... which means that I'll have to retest everything out to find the best settings, right?

[violet coral]

Ah alright

[grim geyser]

wdym sounds more distorted?

[violet coral]

Uhh

It sounds as if the speaker was grinding through fine sandpaper

Mixed in with the voice

[grim geyser]

so, little noise and robotic?

[violet coral]

Yep

Don't mind the analogy

[grim geyser]

maybe try to see if this is related to specific models

you can't use sup2 since it's on client only

[violet coral]

A

[grim geyser]

did this issue come out only after lowering the extra? or was this on some specific models?

[violet coral]

Just one specific model

The other ones work relatively better

As in less robotic

[grim geyser]

it's a model issue then

not all models are high quality

they are made by the community and there's over thousands

[violet coral]

That's true

The model that has issues was the Alya Kujou one posted in this server

[grim geyser]

it's better to search for other models

[violet coral]

This one

#1268088600415830078 message

[grim geyser]

the only way is to either test them urself or listen to the sample

[violet coral]

Ah alright

So in essence

To fine tune the models

Decreasing Chunk Sizes results in better performance as long as it's within a 30-50 upper quartile range from the perf

Decreasing Extra results in faster responses but slightly worse-for-wear voices

Increasing GPU usage leads to better sounding voices but at the cost of delay

Right?

[grim geyser]

can't listen right now but @spiral flax said that model works better with softer voices #1268088600415830078 message maybe it's the reason why only that specific model gives you a bit of noise ?

did this model also give issue before changing the settings?

[violet coral]

It didn't

I've been using the same level of my voice in testing to make sure that it didn't factor into the changes

[grim geyser]

chunk controls delay
be sure it's like 60ms over than the perf number

[grim geyser]

extra controls a bit quality
lower values mean better performance, less risk of cutting off issues, but can cause worse quality

[grim geyser]

so, this model works better with an extra like 2.7 but that causes you to have cutoff issues?

[violet coral]

Yes

Ok so setting chunk to 122.7ms and Extra to 2s results in a better audio quality for me

[grim geyser]

and does this specific model still have issues on those settings?

[violet coral]

Yes

I’ll probably just not use that specific model

Every other model works perfectly fine

None of the models does very well with sudden inputs such as coughs or sneezes but I don’t think there’s much that can be done about that

[grim geyser]

if you'd still like to play with it, you might try setting extra to 2.7 and maybe speaking nearer to the microphone or even trying noise suppression in client mode

not sure if @winter wedge got any other ways

[grim geyser]

RVC has limitations on things like laughing,
coughs and sneezes might be ones too

-# i mean tbf sneezing with an ai model isn't super important

@violet coral do you need any other help or is this solved ?

[violet coral]

Should be good I think

I’ll play around with it a bit more later on

But this is a good starting point

Thanks for your help!

[grim geyser]

alright, will mark this as solved then, for future issues make a new post

you're welcome

[winter wedge]

Your gpu is going idle, so lowering chunk (more strain on gpu) makes it more active

Try running the force_gpu_clocks.bat outside of the mmvcserversio folder

Should fix the cutting out at the beginning (because that way its constantly running without going idle after prolongth seconds of not speaking)

[violet coral]

What should I do if this is what it says whenever I try to run it?

[limber sail]

Two batch files outside the MMVCServerSIO folder are "nvidia-smi" commands that one locks your GPU clock speed while another one frees your GPU clock speed. nvidia-smi is a utility program bundled with every NVIDIA GPU driver. However, any attempt to run either these two batch files without "run as administrator" will result in these errors.

[violet coral]

Ah alright

I’ll try running it later today and message back later

Also would you recommend this (https://github.com/w-okada/voice-changer) or this (https://huggingface.co/Shadicti/deiteris-Fork/blob/main/voice-changer-windows-nvidia-b2332.zip)?

I’m using the second one but was wondering if the first one showed better results or not

[limber sail]

I'd rather recommend the second link. The first link is the "original" W-Okada.

[violet coral]

kk

I’m using the second one currently because I thought it was a more optimized version of it

[limber sail]

The "second link" is Detris' fork W-Okada, and it's a more optimized of original W-Okada. Also, there's a GitHub for the second link. https://github.com/deiteris/voice-changer

[violet coral]

Ah alright, thanks!

Also I can’t check right now so I’m asking, but is the second version a PWA by any chance?

It always opens up a new tab on my browser and I was wondering if there was a way to make it a standalone icon(?)/window

[acoustic iron]

see the guide https://rentry.co/forkvoicechangerguide

[violet coral]

Short answer: No, because it’s a WebUI and helps with optimization compared to running it on a standalone window

[limber sail]

To make fork W-Okada to open as a separate window, download the codebase from GitHub and learn how to make GUI for that.

[violet coral]

kk

Doesn’t sound too hard :D

I’ve done worse

Way worse before

[limber sail]

If you're unaware of what you're doing, better leave it as the way it is.

[violet coral]

Eh I’m aware of what I’m doing

To a certain extent

But I’ll honestly probably keep it as it is because of exams coming up soon

And just do it as a quick summer project

Ah right

Also for the gpu clock program, will it reset automatically or will I have to change it by running the reset bat?

And does it have any adverse effects-?

[acoustic iron]

the standalone window in the og wokada is actually a kind of Chrome (CEF) framework, and the intensive tasks are handled in the python backend (the console window), not the UI frontend

so your argument is invalid

[violet coral]

“Because it uses a Web User Interface (WebUI) coded in JavaScript & TypeScript, the majority of AI programs are designed to run on the browser since it can be used both on cloud and locally (on your pc), and it can also resolve some performance issues. The original wokada also ran on a WebUI, just that it made it's own window” is what the guide said 😭

Also forgot to add a question mark there

Was asking if that was right

TwT

[grim geyser]

it's better not, it uses a web user interface as the majority of the ai programs, it's because it can be edited easily and also can be used on the cloud

the original wokada uses the web user interface too but on it's own window, and that caused it to have worse performance and some cases even crashed

just use the browser

[violet coral]

kk

[grim geyser]

so this is solved, right?

[violet coral]

I hope so lol

You can mark it as resolved; I’ll open up a new ticket if there are any future problems. Thanks!