So there was some discussion on using AI to create hypno files, which inspired me to trial the technology out there and create something viable that is not just mainly luckshots and slot machining like 3 minute long clips. So what we really need is people providing transparent demos maxxing on the tech. And people vouching that the results are not based on slot-machining and getting random outcomes. So people can create actual hypnotist personas + scripts that are consistent to what they specified exactly, not whatever random thing the generator threw in their face. Also it would be massively important to mention how much time you spent.
With Dramabox you can script how things are said and change intonation and voice style etc. And it also does breathing sounds, whispering, gasping, tons of novel inclusions of non-speech elements in the training data that older TTS just were not capable of. But it does not do porn moans, albeit the whispering sounds decently erotic and sensual in my opinion.
In the end I thought a fun test would be to make a "companion" audio to an ASMR video.
You can listen to the result here: https://gofile.io/d/GovtMojx
You are supposed to play this audio simultaneously and a little louder than the ASMR video it comments on: Gibi ASMR | Devilish Triggers
!!!PLAY BOTH LINKS SIMULTANEOUSLY!!!
It was a button-press variation of a general concept I fleshed out over like 6 hours before, on how hypnosis files are structured. This concept is not perfect and just one day of work, but it works for many future purposes and the overhead to make new files is minimal by comparison. You need to add more stuff to it if you want to create more files, so it doesn't sound repetitive.
This is the concept's summary: dpaste.org/gXt2n/raw
The concept's example result: dpaste.org/1ahSL/raw
I made this concept with sample from same speaker, and found that other speakers did work. But also that the concept was sort of grown organically around the nature of the speaker's voice (ASMR soft/whispered/deep). So some adaptation might be required, if your speaker is like "bratty teen" or "dominant cugar". I then later sampled 30 seconds total from various places of an undisclosed ASMR video (not from Gibi obviously for legal reasons ...) and it sounded surprisingly like Gibi in the end. I harvested 3-5 times, and checked each clip compilation how well it worked for the purpose. And there is a remarkable difference sometimes with certain moods and voice styles that you can't foresee. I also tried actually sampling Gibi videos for the TTS (just for testing, not for publishing), but I noticed that the results sounded somewhat different to Gibi and it creates this unique identity of speaker still. So this would also be relevant for legal reasons, that you can't 100% ID the sample origin and it is like very transformative in a legal sense but still not generic.
The sample harvesting, it takes maybe 10 minutes per sample and then a couple more to try it out in Dramabox. Doing this cleanly and properly is most important and you should not cut corners here and put in a lot of effort to understand what you are doing and the TTS is doing.
So to get the script in the video, I just prompted Deepseek to use this concept and the example result and transform it to make it work to comment on the video. Like it was a 5 Minute effort, and then the timestamps misaligned and the content was not how I liked. So again 2 times I told it make this longer that longer. 15 Minutes the script was done.
Here is the script of applying concept + example to yield new script: dpaste.org/iqZTq/raw
To finalize the audio, I used some hack-and-slash script from ChatGPT that fills in the stupid 500ms gaps of silence that Dramabox creates when it stitches the 45s long segments together. This segmentation is also a major reason why you should only prompt this with another LLM and never write scripts by hand. The LLM will (or at least should according to concept summary) bloat the script by repeating all voice set and settings over and over each line. If it does not do that your LLM is shit.
Here is the silence-gap-fixing python script: dpaste.org/o0Jxh/raw
Now at this point, the file itself sounds reasonably well on its own - but really you need second tracks like the ASMR video it comments on, so it becomes really good. If you did a standalone audio, the other tracks could be the binaural beats from hypnosis, plus a layered background track with instructions and/or just generic ASMR like lotion sounds. It would be some more hours of effort. But since it was commentary to the Gibivideo this was not needed.
However ... I still put in like 2 more hours or so because I repaired certain triggers that were misspoken by the TTS, sounded off and the segments did not align perfectly and it didn't follow instructions to not say "she is picking up candle now", because how the fuck is the LLM supposed to know from timestamps what and where she actually does things ... whatever then it was late and fuck this shit.
So I found it pretty impressive, you can use this for all kinds of purposes. This was just a shower-though thing, to make sexualized hypnosis commentary on ASMR videos. If someone spents like 20 hours straight to make this concept more universal with variation, you could also listen to it in the background and it encourages you while working or taking a dump or whatever. I haven't really thought about all the possibilities here. That's something I leave to you.
So since I already did all this work, I am going to estimate how much time it takes for you to create another audio of same quality:
* Install WSL2 (Linux inside Windows) with CUDA etc: 2 Hours
* install Dramabox: 10 minutes + 16GB(?) download
* Start Dramabox: 3-5 Minutes
* write like 1000 words to personalize your concept with Deepseek or local LLM: 15 minutes
* generate audio at 3-5x speed: 5 Minutes
* write 500-100 words again to LLM to refine the outcome: 10 Minutes
* repeat last 2 steps multiple times
* add more audio layers for A+ result: 2-5 hours?
Requirements:
* 24GB GPU memory
* CUDA(?)
The settings for Dramabox are very simple:
* STG Scale: try 3 to begin with or higher for experiment
* Duration: About 1.4-1.5 for hypnosis and something like 1.1-1.2 would be very fast, like for a climax segment you can generate separately
* Reference Duration: 30s to get max audio sample
Here is the full Deepseek chatlog I had this day for any more details: chat deepseek.com/share/bbdd4orgftol3hgu52
I hope this helps you to get an overview of how much time this shit needs and where the limits are of the tech and what you can do with it.
With Dramabox you can script how things are said and change intonation and voice style etc. And it also does breathing sounds, whispering, gasping, tons of novel inclusions of non-speech elements in the training data that older TTS just were not capable of. But it does not do porn moans, albeit the whispering sounds decently erotic and sensual in my opinion.
In the end I thought a fun test would be to make a "companion" audio to an ASMR video.
You can listen to the result here: https://gofile.io/d/GovtMojx
You are supposed to play this audio simultaneously and a little louder than the ASMR video it comments on: Gibi ASMR | Devilish Triggers
!!!PLAY BOTH LINKS SIMULTANEOUSLY!!!
It was a button-press variation of a general concept I fleshed out over like 6 hours before, on how hypnosis files are structured. This concept is not perfect and just one day of work, but it works for many future purposes and the overhead to make new files is minimal by comparison. You need to add more stuff to it if you want to create more files, so it doesn't sound repetitive.
This is the concept's summary: dpaste.org/gXt2n/raw
The concept's example result: dpaste.org/1ahSL/raw
I made this concept with sample from same speaker, and found that other speakers did work. But also that the concept was sort of grown organically around the nature of the speaker's voice (ASMR soft/whispered/deep). So some adaptation might be required, if your speaker is like "bratty teen" or "dominant cugar". I then later sampled 30 seconds total from various places of an undisclosed ASMR video (not from Gibi obviously for legal reasons ...) and it sounded surprisingly like Gibi in the end. I harvested 3-5 times, and checked each clip compilation how well it worked for the purpose. And there is a remarkable difference sometimes with certain moods and voice styles that you can't foresee. I also tried actually sampling Gibi videos for the TTS (just for testing, not for publishing), but I noticed that the results sounded somewhat different to Gibi and it creates this unique identity of speaker still. So this would also be relevant for legal reasons, that you can't 100% ID the sample origin and it is like very transformative in a legal sense but still not generic.
The sample harvesting, it takes maybe 10 minutes per sample and then a couple more to try it out in Dramabox. Doing this cleanly and properly is most important and you should not cut corners here and put in a lot of effort to understand what you are doing and the TTS is doing.
So to get the script in the video, I just prompted Deepseek to use this concept and the example result and transform it to make it work to comment on the video. Like it was a 5 Minute effort, and then the timestamps misaligned and the content was not how I liked. So again 2 times I told it make this longer that longer. 15 Minutes the script was done.
Here is the script of applying concept + example to yield new script: dpaste.org/iqZTq/raw
To finalize the audio, I used some hack-and-slash script from ChatGPT that fills in the stupid 500ms gaps of silence that Dramabox creates when it stitches the 45s long segments together. This segmentation is also a major reason why you should only prompt this with another LLM and never write scripts by hand. The LLM will (or at least should according to concept summary) bloat the script by repeating all voice set and settings over and over each line. If it does not do that your LLM is shit.
Here is the silence-gap-fixing python script: dpaste.org/o0Jxh/raw
Now at this point, the file itself sounds reasonably well on its own - but really you need second tracks like the ASMR video it comments on, so it becomes really good. If you did a standalone audio, the other tracks could be the binaural beats from hypnosis, plus a layered background track with instructions and/or just generic ASMR like lotion sounds. It would be some more hours of effort. But since it was commentary to the Gibivideo this was not needed.
However ... I still put in like 2 more hours or so because I repaired certain triggers that were misspoken by the TTS, sounded off and the segments did not align perfectly and it didn't follow instructions to not say "she is picking up candle now", because how the fuck is the LLM supposed to know from timestamps what and where she actually does things ... whatever then it was late and fuck this shit.
So I found it pretty impressive, you can use this for all kinds of purposes. This was just a shower-though thing, to make sexualized hypnosis commentary on ASMR videos. If someone spents like 20 hours straight to make this concept more universal with variation, you could also listen to it in the background and it encourages you while working or taking a dump or whatever. I haven't really thought about all the possibilities here. That's something I leave to you.
So since I already did all this work, I am going to estimate how much time it takes for you to create another audio of same quality:
* Install WSL2 (Linux inside Windows) with CUDA etc: 2 Hours
* install Dramabox: 10 minutes + 16GB(?) download
* Start Dramabox: 3-5 Minutes
* write like 1000 words to personalize your concept with Deepseek or local LLM: 15 minutes
* generate audio at 3-5x speed: 5 Minutes
* write 500-100 words again to LLM to refine the outcome: 10 Minutes
* repeat last 2 steps multiple times
* add more audio layers for A+ result: 2-5 hours?
Requirements:
* 24GB GPU memory
* CUDA(?)
The settings for Dramabox are very simple:
* STG Scale: try 3 to begin with or higher for experiment
* Duration: About 1.4-1.5 for hypnosis and something like 1.1-1.2 would be very fast, like for a climax segment you can generate separately
* Reference Duration: 30s to get max audio sample
Here is the full Deepseek chatlog I had this day for any more details: chat deepseek.com/share/bbdd4orgftol3hgu52
I hope this helps you to get an overview of how much time this shit needs and where the limits are of the tech and what you can do with it.