[ISSUE for v2]: and v1. Words like "chat" are very deep and smoker voice ish

[FlossDude]

Voice Changer Version

Multiple versions of v1 and v2

Operational System

Windows 11

GPU

Nivida

CUDA Version

Don't know how to find out the cuda version?

Read carefully and check the options

• If you use win_cuda_torch_cuda edition, setup cuda? see here
• If you use win_cuda edition, setup cuda and cudnn? see here
• If you use mac edition, client is not launched automatically. Use chrome to open application.?
• I've tried to change the Chunk Size
• I've tried to set the Index to zero
• I've read the tutorial
• I've tried to extract to another folder (or re-extract) the .zip file

Does pre-installed model work?

No

Model Type

RVC

Issue Description

Hello. It's been a few months, I didn't report this issue sooner as I figured it might just be a issue with the version I used as I was originally using a older version of v1. However even update v1 and v2 has the same issue, no matter what voice model I use certain words almost always come out as a deep robotic/smoker voice sound such as chat, back anything with at or ack sound in.

At first I assumed it was just the rvc model I used so I switched to another and same issue and another and another. Tried v2 same issue on all the same models. Even tried converting them to ONX and still the issue. Even the infamous Amanda Silva voice has the issue so I'm almost certain now it's not the models.

Is this a common issue? And if so is there a fix for it?

As for settings I've tried so many different ones over the months i can't list all I tried. But I can list my current settings on both v1 and v2 if required ^^ even different pitches high

And versions I've used a few different versions. I'm currently not at my pc so can't get the exact versions in current using. But I know for a fact I'm not using the newest v2 that apparently came out yesterday or a few days ago. I only just saw in another post that there's apparently a new version came out. I'll download it when back at pc.

Also this is my first post so sorry if I'm lacking the information required I'm on mobile atm at work. Thanks for the patience and low.

Application Screenshot

No response

Logs on console

Currently on mobile. Let me know how to check the log on PC and will update this when home.

[FlossDude]

vcclient_win_cuda_2.1.4-alpha
And
MMVCServerSIO_win_oonnxgpu-cuda_v.1.5.3.18a
Are the versions I use. I even tried the forked version

My GPU is Navida GeForce RTX 4070.

Current Settings for v2 are
Input: 1.0
Output 1 and 2: 1.0
Pitch Estimatior: rmvpe
Pitch: 7
Index: 0
Chunksec: 24000[0.5 secs]
Extra: 480000[10 secs]
Embedder: Default hubert base 112

v1 current settings are
In: 1
Out: 1
Tune: 7
Index: 0
Chunk: 768
Extra: 131072
F0 Det: rmvpe-onx
Protocol: sio
Crossfade: 4096, start 0.1 end 0.9
Trancate: 300
Silentfont: On
Protect: 0.5
RVC quality low (sometimes high sounds good other times it lags and sounds robotic, dosn't change the low smoker ack sounds)
Skip pass: no

[Kuuko-fokkusugaru]

To be honest, your first assumption was right, it definitely depends on the model. It doesn't matter if you think that a certain model should be better; most of them are not that good, really. The thing with models is complicated; the quality can't be easily measured. And there are a lot of different factors too, like your accent. Basically, most models are trained with very little audio data. Often, it's just a few seconds instead of a proper clean audio with enough words to capture all the phonemes. That said, you also need to account that every model has a main language of origin. I have never found any issues pronouncing most of the words, and it is only with other noises that may fail. In fact, I have certain voice models that can do laugh, whisper, sneeze, cough, etc, just fine (though, this is not the norm, sadly).

So, as you can guess, the included models are mostly made for Japanese and similar language phonemes. English may not be the best for it, as the training data will lack the phonemes needed for certain words. It doesn't mean that you can't use them; it means that it may not sound exactly as you expect. And you may think, "But there are also English voice models that still sound bad". And that's right, and it happens because a lot of people use English audio to train the model, while not actually training it set to English. I have come across models trained in a Japanese base but using English audio, and the end result is lacking.

That said, there are other factors too. The pitch is probably the main one to sound robotic. Here is how it works. Unlike what most people think, the pitch is not meant to configure the voice to your taste. Instead, it is meant to be the pitch of "your voice". You need to adjust the pitch to make sure that it sounds closer to the original voice (as long as you know who the original voice from or you have clips of audio from the original to listen to. The closer to the original, the more realistic and better it will sound. This is easily tested when you leave it at 0, and you sound completely robotic. But, funny enough, when I tried certain voices going from girl to girl, I just had to set the pitch to 1 or 2. That's because we may have a similar tone. If you go above or below the sweet spot, you can get distortion and a robotic voice. You can also get it if, while speaking, you tend to naturally change your voice pitch. I have seen this especially in some guys who tend to end the sentences with an extremely forced deep voice or a hoarse one (RVC HATES hoarse voices lol).

So, right now, your best bet is to first get a proper pitch set in the software. Speak naturally and try to mimic a certain sentence from the original "character"; that's the easiest way to know if you are doing it right. Just don't try to make their "voice", speak with your natural voice. Avoid sounding weird, like with an extremely deep, raspy, and hoarse voice, because RVC interprets that as "noise". Find a good model, or try to use one that you have heard before. For example, if someone is talking about a certain model and distributing it on YouTube, it's easy to listen to how it sounds for them, so you know how it "should" sound for you. But keep in mind that every single model is, in some or other way, affected by how we speak, so the same model won't always sound the same to different people. My English is far from the best, and my pronunciation is quite bad, but my friend uses RVC too, and her English is good enough, and we have never found many things sounding weird beyond the common ones. So I truly think that this is not really an issue with the software itself, but in that you need to learn a bit to control your voice a bit more and find a good model so it sounds homogeneous and without pitch changes.

If you still can't figure out how to fix that, you can always send some clips over here so we can take a look (or listen lol) to them and see how they exactly sound and a possible solution. BTW, the default truncate setting for v1 is 100, not 300. I have no idea how this can affect the resulting voice, as there aren't many explanations about those settings, and I never messed with them, so keep it in mind in case it may be affecting the result.

[FlossDude]

So, right now, your best bet is to first get a proper pitch set in the software. Speak naturally and try to mimic a certain sentence from the original "character"; that's the easiest way to know if you are doing it right. Just don't try to make their "voice", speak with your natural voice. Avoid sounding weird, like with an extremely deep, raspy, and hoarse voice, because RVC interprets that as "noise". Find a good model, or try to use one that you have heard before. For example, if someone is talking about a certain model and distributing it on YouTube, it's easy to listen to how it sounds for them, so you know how it "should" sound for you. But keep in mind that every single model is, in some or other way, affected by how we speak, so the same model won't always sound the same to different people. My English is far from the best, and my pronunciation is quite bad, but my friend uses RVC too, and her English is good enough, and we have never found many things sounding weird beyond the common ones. So I truly think that this is not really an issue with the software itself, but in that you need to learn a bit to control your voice a bit more and find a good model so it sounds homogeneous and without pitch changes.

If you still can't figure out how to fix that, you can always send some clips over here so we can take a look (or listen lol) to them and see how they exactly sound and a possible solution. BTW, the default truncate setting for v1 is 100, not 300. I have no idea how this can affect the resulting voice, as there aren't many explanations about those settings, and I never messed with them, so keep it in mind in case it may be affecting the result.

The models I use are English trained models. I never liked the quality of the default models, they sound like vocalids, which I guess. Makes sense since W-okada was orginaly for singing.

But even most the English models I can find like the Amanda one has like a weird, I guess reverb? The reverb is significantly reduced by removing the index file or setting it to 0. So the model I find to be the best quality is one I've used the most, can't even remember where I got it but it's just called "Female" and it has no reverb it dosn't even have a index file just a PTH and it's the one that matches my real voice the most, only issue is again the very harsh deep smoker voice sound it gives when saying ack words. But all models I've tried had this issue.

I don't like gong higher than 7 on the pitch as it becomes less real sounding and more chipmunky the higher it gets. But still has the deep smoker acks even at higher pitches. I haven't tried Lower pitches but lower pitches.

And yeah I set the truncade to 300 as it was what was recommend in a video I saw a while back. Tbh I haven't heard any difference with it at 100 or 300 but apparently it's supposed to help It sound less robotic, but robotic can mean a lot of different things. I would count the deep smoker ack sounds as robotic, but I also count like the reverb of the default models as robotic to.

[Kuuko-fokkusugaru]

Yeah, but even then, some of those English models are trained with English voice clips but with a Japanese base. Basically, when you train the model, the software to generate the voice models expects certain language and phonemes. More often than not, people don't change the default settings and end up using English voice clips to train a Japanese model. Obviously, it will still work, but the software it's mapping English phonemes into Japanese ones, which aren't the same, and gives bad results. Not saying that it's the case with your model, but it happens too often.

The Index file it's, let's say, the "accent". When you use an index file, the software will search it for the different phonemes, so it tries to match how the original character said it. This does not actually affect the quality of the voice; what it does is try to mimic how that person speaks, so you sound the same. For example, if you set the index to the max on an English voice and you try to speak a language with a very different set of phonemes, like Spanish, it will sound like an English person trying to speak Spanish with an English accent, no matter how good you are at pronouncing Spanish, lol, and it's funny. But if the voice is English and a Spanish person tries to speak English with a maxed index file, this person will get aid on actually sounding more English and better. Most of the time, unless you are trying to impersonate a character, you may want to leave the index at zero.

The reverb thingy happens often because the original audio used to create the model does have reverb. This happens often when the audio isn't clean enough and has been taken from a speech or a stream where the person originally had reverb. Ideally, the source voice clips would be cleaned of noise and reverb before training the model, but not everyone does this, and sometimes it is also quite hard to do. There are really many models with reverb and even noise sounds in the background due to the same thing.

Going back to the pitch, like I said, isn't a matter of setting it lower or higher. It's about setting it to sound with the exact same tone as the original voice (or as close as possible). Let's say the original voice is actually quite high-pitched and like a chipmunk, like you said, then you will need to actually set it high enough to achieve the same chipmunk voice; otherwise, it won't sound as good, and you will be more prone to robotic and distortion effects. Of course, this is just an example, but you really need to match the original tone. You can change the pitch by processing the audio with a different software later on, or through some of those non-AI voice changers that let you change pitch in the old way.

And going back to the wrong results of words ending in "ack", you already know the answer from my previous reply. The way you pronounce those words may actually lower your pitch naturally or make your voice sound more raspy. Regardless of the reason, it's not the word itself but how you pronounce it. Your best bet is to train those words and try to keep the same tone and pronounce them very clearly. You can also make sure that your mic is not already applying any noise reduction or filter that may actually be taking off quality to the end of those words by suppressing details. You can check this easily by pronouncing these words differently. Try changing your tone on purpose for the sake of testing. If you are a guy, soften your voice and raise your pitch to sound like a girl, and try to say those words in a clearer way, making sure that you are speaking loudly enough. Then you can test if they actually sound better and not robotic.

There is really no other way than practicing those words with your normal voice, but trying to sound clearer. Think of it as if you were voice acting a bit.

[FlossDude]

Thing is idk what the orginal voice is supposed to sound like. It's just called Female 😅 so my guess is it shouldn't be set to high and chipmunky. But a middle like 7 or even lower.

And yeah I like to avoid the revered sounding models. Sadly I suck at finding good ones. Only good one I have is Female which has no reverb ect.

I've tried about 9 different models all but female has a reverb and all the harsh smoker on ack. If it's my voice naturally lowering when I say words like chat and words with ack in then testing other models wouldn't help?. But you mentioned you had some good English models you've tried. Do you mind if I try them to?.

[Kuuko-fokkusugaru]

Well, there is not a good pitch value that works for all. It all depends on your voice. From male to female it sits usually at around 12 but this also depends on how high or low your voice is. Someone with a light teen kind of voice may be fine with something between 8 or 10 while someone with a really deep voice may need 15 or more. This also depends on the target voice model; the further the difference between its tone and yours, the further the pitch numbers may go.

Sadly, the models I have are paid ones and they aren't English, they are Japanese ones but it seems to work fine with English (or more like my Engrish XD). Maybe with a native person it may not sound that great. The model that my friend uses also works though, it's a free model that she got from the some website or the AI Discord I belive. I think it's from a vtuber named Nina Momose or something similar (I know that the last name is Momose 100% but can't remember the name).

You could try to upload some audio clips so I can see what exactly do you mean by robotic. And if you aren't shy of showing your own voice, you could send clips of you speaking and saying the words that gives those bad results. Just make sure to record a whole sentence or a couple of them, don't just send the word "back" sounding bad or I wouldn't be able to compare between good and bad sounding lol. I could then run your clips using the file option in RVC and test some of the voices I have and see if the result is the same as yours or if it sounds better. And if it sounds better then we could go further and I could test those same clips with the same free models that you have and see if I am getting the same results that you do or if in my environment it sounds better. In that case we would know for sure if it may be a software issue that causes RVC to bug and give bad results to you. But if we get the same result then we would know that is all about pronunciation plus software limitations at time of understanding different voices, tones, and accents.

Keep in mind that github doesn't allows audio clips to be uploaded for some reason but you can zip them.

[FlossDude]

I'll be honest I don't know how to zip a file 😅 only how to unzip. And my voice is kinda unique, English but I don't sound like your stereotype. So even when using a voice module or not my voice is easy to spot. Kinda worried about putting my voice on the Internet raw since I've already worked on projects with people that are out on the Internet with my current module.

[FlossDude]

Also. If it is the moduls that cause the deep harsh sounds how worth it would you recommend buying paid modules and where would I find them? I've only been able to find free ones. And since you said your paid modules have no issues.

[Kuuko-fokkusugaru]

Remember that I said that I have some models that sounds great and they catch many details of my voice and that they just happen to be paid ones. But this doesn't means that all paid models are great. I have many that aren't capable of doing so and they are still paid ones. I often get them from booth but my comment wasn't intended to imply that only paid ones may be good ones. Also, I mentioned that my friend uses free ones and they work fine for her. So, while the models have a huge impact on the quality, we don't know if this is a critical factor to fix your issue since I couldn't hear your robotic back examples nor your original voice. I am just making assumptions and telling you my own experience with the software with my own voice and that of my friends.

[FlossDude]

Remember that I said that I have some models that sounds great and they catch many details of my voice and that they just happen to be paid ones. But this doesn't means that all paid models are great. I have many that aren't capable of doing so and they are still paid ones. I often get them from booth but my comment wasn't intended to imply that only paid ones may be good ones. Also, I mentioned that my friend uses free ones and they work fine for her. So, while the models have a huge impact on the quality, we don't know if this is a critical factor to fix your issue since I couldn't hear your robotic back examples nor your original voice. I am just making assumptions and telling you my own experience with the software with my own voice and that of my friends.

It's not just back. Back and chat were an example. I actually took part in a project last night so have even fresher in my mind. It seems to be pretty much any word with a in even at seems to have a ack sound. But the overall quality last night was very robotic sounding as a whole tbh. I don't even know why, I might need to re-install v2 since that's the one I used last night.

Also just to confirm I'm using both the most recent versions of v1 and v2? I did try the new v2 one that was uploaded a few days ago but it's Beatrice only so wasn't able to upload my moudels.

[FlossDude]

Okay. So I'm starting to think maybe it's a hardware issues with my Machine?. Since no matter what, I've brought a bunch of payed models on Booth, Ko-fi, ect and tried free models and all have the exact same issue.

The best result I have is on the W-okada fork instead of the official versions. But the issue is still present even then.

I really don't think it's a issue with the pitch as when I set it up to 12+ every model sounds like horrible super high pitched chipmunk no matter what model it is. And the issue still happens even then. All models sound the most realistic at pitch 5-7 even then sometimes 7-8 is pushing it a bit. And regardless of the pitch the deep smoker/robotic sounds still appear.

These don't seem to appear on all the example videos people who sell the models have on booth/ko-fi ect. It's very puzzling to say the least.

[Kuuko-fokkusugaru]

There is no reason to chose 12 or more if your voice isn't deep. As long as it sounds with the same tone as the sample, it should be fine. Like already mentioned, this is meant to fit that of your voice to the model one. About the robotic voice, we hit a wall long ago. I don't know how it sounds and you had your reasons to not provide examples so things won't move forward from there. But no, isn't a hardware issue unless your mic sounds really bad. You can always try with some random audio clip from someone else and compare.