Question Local LLM For Voice Over With Exertion Modulation

[kim05]

Have anyone experimented with any LLMs with emotion control that can be run locally with no censorship? I really want to make a scene with a female trying to talk normally whist getting banged, so that would entail a relatively normal script for the machine to read, but relatively extensive emotion control like gasps and other inflections of exertion. So far Qwen3 TTS (under 8Gb VRAM) and IndexTTS2 had made me feel like I've wasted my time installing them. I'd be delighted if someone could prove me wrong and explain how it's done with those models or advise something else.

Spoiler

Please, leave comments like "AI = bad" on reddit, I'm aware.

 

[atani]

Yeah, man. You have Voxta and some plugins out there, but running locally also means you have the HW power to do VAM and LLMs.

 

[kim05]

Yeah, man. You have Voxta and some plugins out there, but running locally also means you have the HW power to do VAM and LLMs.

No, I mean an LLM with a audio file output. I don't need chatbot with VaM integration, those still sound very much like it wants to sell me some sort of corpoproduct or a subscription, so Voxta is not really for me. What I thought of is more like a triggered voice line type use. Line 1 on Pose 1, Line 2 and 3 on Pose 2 etc. I was surprised by how Qwen3TTS can clone voices so they don't sound robotic, but it cannot do emotion modulation and breath/exertion sounds to that or at least I haven't succeeded at getting such a result. I was curious if anyone tried dealing with this kind of generation for the purposes of AI simply voicing the script with emotion and heavy breathing, not interacting with me directly.

 

[atani]

There's some plugins that do X or Y things, and vamx also has some sort of AI features. I don't know what they do or not do.

For my use I do AI cloned voice audio sentences files and play them on my scenes during a story intro or more commonly during the saucy sex animations. They have some emotion depending on the source material and the model used, currently Index-TTS and Higgs, but I create them all beforehand, not in real time with "script interpretation".

 

[rokoc]

I had some limited success with coqui xtts2 cloning locally and using samples of audio from solo porn vids, asmr audio pron, etc.

That's probably a bit dated these days though and something better might now be out but I've not really bothered looking in to it again. it was mostly out of curiosity.

Impressive results to say the least but In the end it was too much work involved creating a good enough dialogue

 

[atani]

If you're looking for better and up to date tools to create AI voices or AI cloned voices audio files, look no further:

GitHub - diodiogod/TTS-Audio-Suite: A ComfyUI custom node integration for local multi-engine multi-language Text-to-Speech and Voice Conversion. Supports: RVC, Echo-TTS, Qwen3-TTS, Cozy Voice 3, Step Audio EditX, IndexTTS-2, Chatterbox (classic and m

A ComfyUI custom node integration for local multi-engine multi-language Text-to-Speech and Voice Conversion. Supports: RVC, Echo-TTS, Qwen3-TTS, Cozy Voice 3, Step Audio EditX, IndexTTS-2, Chatterb...

github.com

11labs can suck it.

 

[hazmhox]

Even the most advanced cloning services can't pull that off. Highly unlikely you'd find a free open source one able to do that properly.

 

[Clownmug]

Rather than going the TTS route, why not put your script through video+audio gen (with nsfw LoRA) like LTX or MiniMax H3? The speech might be a little more natural when driven by a video.

 

[kim05]

Even the most advanced cloning services can't pull that off. Highly unlikely you'd find a free open source one able to do that properly.

Gemini (aistudio.googledotcom) did all I asked to specs and left me impressed to be honest, when it comes to emotion and inflection. It only has presets for voice though. Qwen3TTS cloning was extremely accurate, but no emotion control for cloned voice. IndexTTS2 has a lot of promise, but I haven't had any good yield from trying since it has higher recommended VRAM requirement than I can afford. I am trying to figure out if there is some formula, a chain of model out there that would permit to finally output something exactly to my specified instruction. I do not hope to find the philosopher's stone, you understand, I'm just looking at the current open-source possibities and pondering the future if anything. And if I will find a sufficiently effective chain or a wonder model than that's even better. That's the point of asking to see the things I haven't yet.

 

[kim05]

If you're looking for better and up to date tools to create AI voices or AI cloned voices audio files, look no further:

GitHub - diodiogod/TTS-Audio-Suite: A ComfyUI custom node integration for local multi-engine multi-language Text-to-Speech and Voice Conversion. Supports: RVC, Echo-TTS, Qwen3-TTS, Cozy Voice 3, Step Audio EditX, IndexTTS-2, Chatterbox (classic and m

A ComfyUI custom node integration for local multi-engine multi-language Text-to-Speech and Voice Conversion. Supports: RVC, Echo-TTS, Qwen3-TTS, Cozy Voice 3, Step Audio EditX, IndexTTS-2, Chatterb...

github.com

11labs can suck it.

This is exactly why I've asked. Wonderful suggestion, I'll investigate with pleasure. Thank you!

 

[kim05]

Rather than going the TTS route, why not put your script through video+audio gen (with nsfw LoRA) like LTX or MiniMax H3? The speech might be a little more natural when driven by a video.

Haven't thought of that, honestly. Perhaps it is worth looking into. My logic was, initially, that I don't need video, so that idea just went over my head completely. Thanks!

 

[kim05]

I had some limited success with coqui xtts2 cloning locally and using samples of audio from solo porn vids, asmr audio pron, etc.

That's probably a bit dated these days though and something better might now be out but I've not really bothered looking in to it again. it was mostly out of curiosity.

Impressive results to say the least but In the end it was too much work involved creating a good enough dialogue

This one I've seen and it does seem like it is outperformed by Qwen3 and IndexTTS2, so I did not try. Maybe worth trying if I'm at my wit's end in the future.

 

[Yosemite Sam]

I'm not exactly sure what you're trying to do here, you want to run a LLM locally, but have the character talk normally while getting banged? Well I mean voxta does that! As for gasping you can just use dragontales body triggers and then put a gasp voice file there for where ever you touch, also the voxta touch elite triggers work pretty good, the AI model will react to different touch zones. You also need to find an actual uncensored model from hugging face, I've probably downloaded about 15 models and if your looking for one that actually moans and gasps well let me know because I haven't found one yet. I do have a uncensored one and it works pretty good with voxtas newest update. Xtts coqui in voxta does a good job of voice cloning if you have a 30 second or longer clip to give it. If you feed it a 10 second clip the voice will be trash, Also with 8gb of vram you are on the low side, I have 12gb and it's low, You will have to configure your models to run on cuda or cpu I think I have two on cpu and one on gpu, I will have to look again, but this configuration is only if you plan on using VR with it. Voxta is not a subscription either you buy it once and then buy it again six months down the line when they update it, which they do very frequently. I dont really want to shill voxta too much but there is nothing out there that compares to it. It does animations, image gen, video blah blah...

 

[rokoc]

I'm not exactly sure what you're trying to do here, you want to run a LLM locally, but have the character talk normally while getting banged? Well I mean voxta does that! As for gasping you can just use dragontales body triggers and then put a gasp voice file there for where ever you touch, also the voxta touch elite triggers work pretty good, the AI model will react to different touch zones. You also need to find an actual uncensored model from hugging face, I've probably downloaded about 15 models and if your looking for one that actually moans and gasps well let me know because I haven't found one yet. I do have a uncensored one and it works pretty good with voxtas newest update. Xtts coqui in voxta does a good job of voice cloning if you have a 30 second or longer clip to give it. If you feed it a 10 second clip the voice will be trash, Also with 8gb of vram you are on the low side, I have 12gb and it's low, You will have to configure your models to run on cuda or cpu I think I have two on cpu and one on gpu, I will have to look again, but this configuration is only if you plan on using VR with it. Voxta is not a subscription either you buy it once and then buy it again six months down the line when they update it, which they do very frequently. I dont really want to shill voxta too much but there is nothing out there that compares to it. It does animations, image gen, video blah blah...

What's hard to understand. They want the girl talking but with all the vocal emotion/tone that sex would naturally give but creating the speech using a local ai program outputting it as an mp3 file to then import to their scene, as a timeline audio track I suspect

 

[kim05]

I'm not exactly sure what you're trying to do here, you want to run a LLM locally, but have the character talk normally while getting banged? Well I mean voxta does that! As for gasping you can just use dragontales body triggers and then put a gasp voice file there for where ever you touch, also the voxta touch elite triggers work pretty good, the AI model will react to different touch zones. You also need to find an actual uncensored model from hugging face, I've probably downloaded about 15 models and if your looking for one that actually moans and gasps well let me know because I haven't found one yet. I do have a uncensored one and it works pretty good with voxtas newest update. Xtts coqui in voxta does a good job of voice cloning if you have a 30 second or longer clip to give it. If you feed it a 10 second clip the voice will be trash, Also with 8gb of vram you are on the low side, I have 12gb and it's low, You will have to configure your models to run on cuda or cpu I think I have two on cpu and one on gpu, I will have to look again, but this configuration is only if you plan on using VR with it. Voxta is not a subscription either you buy it once and then buy it again six months down the line when they update it, which they do very frequently. I dont really want to shill voxta too much but there is nothing out there that compares to it. It does animations, image gen, video blah blah...

Here is my plan:
1. Make a script contextually tied to the scene.
2. Generate AI text-to-speech audio file using that script.
3. Play that audiofile through atom head audio.

I have successfully created a scene like that already. It is up to my specifications, it is near perfect with emotion, delivery, inflection, exertion etc. The issue is I have made it using Gemini that only has a limited number of voices that can be used. It is also limited in amount of tokens and sends everything for analysis to Google. I am not too worried about the contents of my scripts or breach of privacy, but willingly sending shit like "fuck my this and lick my that, mmm so hot bla bla bla" feels a bit like standing naked in Google HQ and reading fanfiction for employees. SO, what I need is an offline model that can:
1. Read a script, preferably with a voice it cloned from reference. Especially useful if I will have an idea for a scene with look-a-like.
2. Modulate it to sound like sexual exertion. Gasps, inhale, exhales, the works.
3. Output it in a singular audio file or a series of audio files that can be later imported into VaM and used with head audio on an atom.

At no point I wish to have an LLM active whilst VaM is booted up and playing a scene. Voxta is something I'm willing to look at later, if I put my mind to different kind of experience in VaM, but from what I've seen it solves a different problem with a different approach, aside from mainly providing services I don't need.

What's hard to understand. They want the girl talking but with all the vocal emotion/tone that sex would naturally give but creating the speech using a local ai program outputting it as an mp3 file to then import to their scene, as a timeline audio track I suspect

Something like that, yes.

 

[Yosemite Sam]

Here is my plan:
1. Make a script contextually tied to the scene.
2. Generate AI text-to-speech audio file using that script.
3. Play that audiofile through atom head audio.

I have successfully created a scene like that already. It is up to my specifications, it is near perfect with emotion, delivery, inflection, exertion etc. The issue is I have made it using Gemini that only has a limited number of voices that can be used. It is also limited in amount of tokens and sends everything for analysis to Google. I am not too worried about the contents of my scripts or breach of privacy, but willingly sending shit like "fuck my this and lick my that, mmm so hot bla bla bla" feels a bit like standing naked in Google HQ and reading fanfiction for employees. SO, what I need is an offline model that can:
1. Read a script, preferably with a voice it cloned from reference. Especially useful if I will have an idea for a scene with look-a-like.
2. Modulate it to sound like sexual exertion. Gasps, inhale, exhales, the works.
3. Output it in a singular audio file or a series of audio files that can be later imported into VaM and used with head audio on an atom.

At no point I wish to have an LLM active whilst VaM is booted up and playing a scene. Voxta is something I'm willing to look at later, if I put my mind to different kind of experience in VaM, but from what I've seen it solves a different problem with a different approach, aside from mainly providing services I don't need.

Something like that, yes.

Right, my answer still stands with voxta running locally on your machine with no internet connection so you can record clips and put them into your scene. Voxta is about $10. I would never be stupid enough to run a LLM on a cloud server lol. Exactly like you said, people reading your chats.

 

[atani]

2. Modulate it to sound like sexual exertion. Gasps, inhale, exhales, the works.

This is by far the most challenging part and will likely sound poor.
I would use VAMMoan sounds for these, even if they don't match the voice of the speech that well.

 

[kim05]

This is by far the most challenging part and will likely sound poor.
I would use VAMMoan sounds for these, even if they don't match the voice of the speech that well.

I can't get around the fact that sometimes some experiments are yielding really good results. It's a thorn in my brain at this point. Like I mentioned, Gemini did a fantastic job at this, but the voice selection and data sharing parts break the deal for me.
Tiny bit of it and I'm not sure I've uploaded it correctly, but here's a sex scene with the female on the phone:

I don't think it's about to win itself some BAFTAs, but damn it's better than all the corporate spokesperson bullshit other models put out.

Qwen3 can do it with one of it's modules, but this module also has a limited premade voice selection and it all sound like a bunch of cheap hentai VAs and even then it only works with some voices.

Overall, this idea of mine is rather old. It was GPT-2 or something like that when I first heard a video of a guy just messing about with it and testing voice modulation at which point it started speaking like someone is sticking a charger into it as it spoke.
With some "free" voice assistants the only input you need to trigger something like that is nonsense like "speak to me as if you're on a jog". Problem is with those assistants, obviously, the refusal rate is quite high even if you try to fool it playing innocent.

I wish I wasn't flat broke at the moment, so I can just hire someone real to do whatever I need, but for now a local LLM is the way to go.
Especially, considering my plan is in a blueprint stage at the moment and I feel like it would be sensible to involve a second party when all else is done and not before.