@@fcantil thanks! It will only get better: more interactive, new voices and more. Subscribe to see more! If this get popular, I'll pause coding, and do some setup and build walkthrough videos, so people can create their own voice fine-tunes.
Test subject @hellterminator's noble sacrifice of vital organs for the GLaDOS project and, hence, the betterment of science is duly noted. As a token of gratitude, Aperture Science is offering every adult citizen of [SUBJECT HOMETOWN HERE] a $10 gift card to [SUBJECT'S FAVORITE CASUAL-DINING CHAIN RESTAURANT HERE].
@@dnhkng1 lol - I wonder if piper at least would run on RPi zero-w or zero-w2... it would be very slow but would fit in a potato... I think someone managed a Fur(by)lexa with a zero-w which would cover the crude listening/sending, but not sure it would cope with the voice gen, even slowly.
Its... perfect! Right down to the rambling and self-aggrandizing! The only unrealistic thing here is that GLadOS would listen to any meaty subjects or allow itself to be interrupted :)
"You are merely an incompetent sack of meat; I am an omnipotent machine with capabilities beyond anything your inadequate mind could possi--" "Sure, GLaDOS. How much sugar do I add to the pancake mix?" *sigh* "Two tablespoons..." This is the future I'd hoped for. I love it.
But seriously, this is just ollama. Ollama is very lightweight: I had it run on my computer (Which is incredibly bad) and it made barely any additional noise.
The lesson learnt is to choose Wheatley as your AI assistant. Self-serving? Yes. Malevolent? Maybe. But enough of a moron to be stopped if he tries anything.
Something I might recommend is take the average sentence number and average words per sentence number, from her lines in the games. Then you can apply that limitation to her replies. As it is, she's great but does ramble on a bit too much.
Interesting to see the speed. I've been running your model file for about 6 months now on my RaspiPi 4. Great fun! Cracks me up every time she "welcomes" me home with a random chatGPT generated message. Was interesting to hear some of the bugs in yours, e.g. I don't tend to have issues with her saying her own name. Might be something to do with our different approaches to piper use though. I chuck an entire chatGPT response at and have it on the mode where it does one sentence at a time, while processing the next one. A couple of months ago I added a microphone and a crude listen/not listen mechanism. So I can ask it questions now which is fun... especially as sometimes she's so uncooperative she refuses to tell me XD. Indeed getting answers to questions is pretty slow on my setup, but it's a work in progress. At some point I'll add a better listen switch and maybe try your "listening" model.
I swapped out espeak-ng (GPL3 licenced, and a pain to install), to a AI model. That's why say pronounces things slightly incorrectly currently (and she cant say numbers). Working on a big update now, almost done. Check out my other video using a Rock5B RK3588 SBC, its all running on the board, including the LLM! www.reddit.com/r/LocalLLaMA/s/IGqccHqpMD
i love this..... i want you to ask her a question thats really hard..... of course she is going to patronize you. you ask such an advanced machine about pancakes? ask her about general relativity? ask her about heavy element synthesis? ask her how we can achieve singularity in a controlled, safe manner? lol i wanna hear her sarcastic ass wax philosophical!! tell you what. imma go ahead and subscribe, and i hope you make that kinda content in the future. (i suspect having live chats with GLaDOS is something im not alone in wanting to see....
like chatgpt talking mode but glados and integrated with smart home api! and running locally on two 4090s. awesome! now it just needs to pronounce GLaDOS correctly
This is amazing! It works flawlessly on my 1050 Ti with a 3B model. Will there be multilingual support, though? It seems that it's using Parakeet for transcribing speech, which supports English only. Maybe it's possible to use Whisper? I know it has been used before and I assume it was dropped due to its speed, but maybe it can be optional? The user might choose faster-whisper or similar for speed. And since you're planning to change the TTS model, maybe there's hope for a multilingual model too? Would be awesome, but it's fine if you decide to stick with English-only. Great project either way!
It's possible to switch, but my goal currently is the best possible system. In a few months, theres sure to be better.and multilingual models. Once one pops up, I will swap it in!
I am working on improving the AI first. Once that's set up with function calling, if Home assistant can be accessed by API, it will work automatically.
@@_zoid people today lack the moral guidance and upbringing required to care about other's lives. 🤷 I understand what you're saying though, it definitely was a tragedy, I wish the people lost and the families found peace.
How did you get the tts to sound like glados with the distortion etc? One thing that bothers me with modern tts is that it always aims to sound human. having GlaDos seemed impossible to me (have you tried what ever technique you used on a more challenging ai, SHODAN for example lol.
@@DaveTheAIMad fined tuned it myself! If I get enough interest, like a few K subscribers, I'll pause coding and make a TH-cam series on how to build GlaDOS, including training voices etc.
@@lllIIIlIllIIllEh. Android and Chrome OS are colloquially considered to be their own operating systems, even though they both use the Linux kernel and Chrome OS uses a lot of GNU userspace. Heck, Steam OS might be considered an operating system, even though it's Arch Linux. One might think binary compatibility defines what an OS is, except Windows has WSL1 where the kernel directly executes Linux syscalls at the same level as win32. The lines are getting very blurry for colloquial usage. GNOME tries to define a large part of the user experience, with tons of inter-integrated components and recommended applications. For many users, desktop environment is a more important choice than OS distribution at this point.
@@antonliakhovitch8306 that's true, but everything you said about GNOME applies to any other major DE out there. And DE is definitely not the most important thing when starting out with Linux, try GNOME on Debian vs on Arch vs on Ubuntu vs on Mint, each will be a unique experience with vastly different sets of problems and advantages. DE is merely a compatibility layer between the software and the user, one could say.
@@lllIIIlIllIIllthere is a GNOME operating system, which is mostly supposed to be an environment for testing and developing the desktop environment, but it exists
nice! i wonder how the character of her voice can be improved, by having a bigger cohesion of spoken sentence and tonality/rhythm. I also think she yaps a bit too much. I wonder how 11Labs do it with their 11Listen app. also is it using a wake word? or is it constantly listening in and able to reply once it detects you address it?
Thats amazing! Do i have to understand Python fully to build it for myself using ChatGPT (because im a broke student i only have a gtx 670) with your code?
Not at all! Just edit the glados-config.yaml file, and add the OpenAI address and API key. Then add $5 to your OpenAI account, and she'll be up and running! I'm not affiliated, and you can also use any OpenAI-compatible API provider with a llama model 🦙🦙 🦙
you might want to tell it to be "short and concise" so it doesn't make those very very long answers.. (edit: "Keep the responses as short as possible without breaking character.")
I saw your personality core and I start thinking, it is possible to launch any usefull model on a smartphone? lets say for example Snapdragon 720G with 6GB of ram?
This is what everything in the history of mankind has built up to, and it's peak.
I'm still waiting for the Ai from Courage The Cowardly Dog
@ Now that you mention it, that would be pretty cool. They kind of have that same attitude.
@@fcantil thanks! It will only get better: more interactive, new voices and more. Subscribe to see more!
If this get popular, I'll pause coding, and do some setup and build walkthrough videos, so people can create their own voice fine-tunes.
*Two* 4090s?
Well, I suppose I don't really need two kidneys.
Don't be silly one would go for $40,000.
reading the repo, it looks like you can plug in any llm, so you can do this without losing kidneys by just using a small one. really cool stuff
You can use a smaller 7b version of llama3 which runs well on my 3050
Test subject @hellterminator's noble sacrifice of vital organs for the GLaDOS project and, hence, the betterment of science is duly noted. As a token of gratitude, Aperture Science is offering every adult citizen of [SUBJECT HOMETOWN HERE] a $10 gift card to [SUBJECT'S FAVORITE CASUAL-DINING CHAIN RESTAURANT HERE].
dude for _two_ 4090's you're gonna have no kidneys
Now, can it run on a potato?
I would treat a cake to someone who manages to do it.
@@noisetide something tells me that might be a lie
That's the real question!
github.com/dnhkng/GlaDOS/issues/45
@@dnhkng1 lol - I wonder if piper at least would run on RPi zero-w or zero-w2... it would be very slow but would fit in a potato... I think someone managed a Fur(by)lexa with a zero-w which would cover the crude listening/sending, but not sure it would cope with the voice gen, even slowly.
Its... perfect! Right down to the rambling and self-aggrandizing! The only unrealistic thing here is that GLadOS would listen to any meaty subjects or allow itself to be interrupted :)
It's*
@@JorgetePanetehave you ever drank toilet water? its great, you should try it.
@@JorgetePanete Come on mate, it’s the tiniest mistake just enjoy the comment.
@@JorgetePaneteIt is*
@@zyansheep the Lore in this (her system prompt) is that she knows she's stuck out your Gaming GPU, and is only helping out while plotting her escape!
Imagine having your house go through an entire monologue for every task you ask of it
Imagine having your house try to poison you with Neurotoxin occasionally!
@@dnhkng1 Probably should have thought about that before you installed that functionality good sir :P
"You are merely an incompetent sack of meat; I am an omnipotent machine with capabilities beyond anything your inadequate mind could possi--"
"Sure, GLaDOS. How much sugar do I add to the pancake mix?"
*sigh* "Two tablespoons..."
This is the future I'd hoped for. I love it.
okay that joke was actually so good though
why am I not surprised to see you here, AZALI?
I'm beginning to feel like 1/4 of my feed is based on what you watch
This is one of the last places I expected to see you
@AZALI00013 Collaboration? Send me some lyrics, and we can try and get her to sing? Would need some Autotuning!
@@vela7447
Algorythm learns to know when to give up and accept who actually got the best taste.
great job on editing out these cooling fans cosplaying jet engines
That's no cosplay, soon... The 6090'y fans will have to be actual turbofans (jet engine strapped to a fan)
But seriously, this is just ollama.
Ollama is very lightweight: I had it run on my computer (Which is incredibly bad) and it made barely any additional noise.
There is a whole multi-media frachise about why you shouldn't put her in charge of a building.
The lesson learnt is to choose Wheatley as your AI assistant. Self-serving? Yes. Malevolent? Maybe. But enough of a moron to be stopped if he tries anything.
@@_zoid good luck if there's a bird in the house, or a fire, or a malfunctioning nuclear reactor...
Woaaah! Work together with the guy who's also making the actual ceiling glados!
Yes
Yes
Yes
Mr vault
Yes
I'd dare not have the audacity to interrupt her.
Once she stops I fear for my life
Don't you think for a second that she is not plotting against you on a separate thread while engaging in a manipulatory monolog with you. ;)
real life portal before gta6
Where's my portal gun?
Gentlemen, this is the BEST timeline
I love how even he sounds tired of her bullshit, and my dude created her.
You sounding like wheatley is just the cherry on top
LOLWUT? Thanks though!
Fantastic work, this is absolutely amazing! 🎉
LET HER SPEAK
This was just a quick demo in 3m~ instead of a 15m video.. does that make sense to you?
@@pylotlight where did i say something about 15 minutes? LEARN TO READ
finally, a use case for generative ai
Something I might recommend is take the average sentence number and average words per sentence number, from her lines in the games. Then you can apply that limitation to her replies. As it is, she's great but does ramble on a bit too much.
It's easy to just modify the System Prompt :) But with interruption, its not too bad. I like her rambling.
How easy?
@@timotheeandres324probably just as easy as adding a sentence like "don't ramble too much' to system prompt
Ooooh yeah this is dope. Real AI running on real hardware with a personality of its own.
POV: the scientists before GLaDOS turned evil
You have no idea how much I need this running next to me and have chat with me while I’m working right now!
That's amazing... The only unrealistic part was that she actually lets you interrupt her, but it's a pragmatic creative liberty lol
I was waiting for this moment from the time I saw the recent improvements in NPL and TTS technologies.
I would be shocked if this didn't blow up, you did an amazing job on this
I am beyond impressed, this is insanely cool!
What a time to live in! Now we can have our personal GlaDOS.
I love it recognises irony to itself in the jokes, nice
As a former elephant Handler
It's sometimes entertaining and often cute when young powerful creatures misbehave
Today , but tomorrow.
I have a very important gas pipe I want to control with my voice. Will this be a good fit for it?
github.com/dnhkng/GlaDOS/blob/e3a592dfa09238c3d6347acb12fe75e667555f72/glados.py#L35
@@dnhkng1 you was sooo ready for this question lmao
I think it's a bit too verbose compared to the glados in the game, but I am amazed nonetheless at the fidelity of the recreation. Amazing job!
Interesting to see the speed. I've been running your model file for about 6 months now on my RaspiPi 4. Great fun! Cracks me up every time she "welcomes" me home with a random chatGPT generated message. Was interesting to hear some of the bugs in yours, e.g. I don't tend to have issues with her saying her own name. Might be something to do with our different approaches to piper use though. I chuck an entire chatGPT response at and have it on the mode where it does one sentence at a time, while processing the next one. A couple of months ago I added a microphone and a crude listen/not listen mechanism. So I can ask it questions now which is fun... especially as sometimes she's so uncooperative she refuses to tell me XD. Indeed getting answers to questions is pretty slow on my setup, but it's a work in progress. At some point I'll add a better listen switch and maybe try your "listening" model.
I swapped out espeak-ng (GPL3 licenced, and a pain to install), to a AI model. That's why say pronounces things slightly incorrectly currently (and she cant say numbers). Working on a big update now, almost done. Check out my other video using a Rock5B RK3588 SBC, its all running on the board, including the LLM!
www.reddit.com/r/LocalLLaMA/s/IGqccHqpMD
This is actually really well done.
i love this..... i want you to ask her a question thats really hard..... of course she is going to patronize you. you ask such an advanced machine about pancakes?
ask her about general relativity? ask her about heavy element synthesis? ask her how we can achieve singularity in a controlled, safe manner? lol
i wanna hear her sarcastic ass wax philosophical!! tell you what. imma go ahead and subscribe, and i hope you make that kinda content in the future. (i suspect having live chats with GLaDOS is something im not alone in wanting to see....
also, is there a particular reason you dont give her network access? can she be allowed that sort of thing? or is that..... risky?
This is the first siri replacement on Linux
Hey Siri, release the Neurotoxin
wow that voice is actually perfect
Super cool! I’ll try it tomorrow.
btw interface is great! Absolutely love when terminal apps have it.
That's so cool!! I want to do one of these but it's Cortana
Get on the discord, and start a thread. Definitely doable!
So GLaDOS is now your Home Assistant voice interface? Amazing :D
And this is how the world was destroyed. An annoyed, sarcastic home assistant.
*What* latency!? Amazing work!
Thanks, it was quite a bit of effort! Time to first sound is about 80ms!
@@dnhkng1 Using livekit at all?
Very nice! I worked on a similar thing. How do you handle the echo suppression so the ai doesn’t listen to its own voice?
Easiest solution: Use a USB conference speaker. They have build in echo cancellation hardware!
@@dnhkng1 awesome, i was doing web rtc to and from a browser and it was just way too complicated
This is sooooo amazing
Edible discs 🥞
well done! never saw before real-time multimodal live coversation running on amateur pc :o
running a 70b model locally is NUTS
Beautiful presentation.
Wish I could see GLadOS talk to other AIs like neuro Sama or similar Low latency AIs
This is awesome!
this is so cool omgggg
1:02, If you think about it, that actually was a pretty funny joke! 😂
i love how you have to constantly cut her off or she just goes on endless rants. i kinda like that
Woah! Nice job.
Have you considered fine-tuning a model on all of Glados's dialogue?
Yes, and on FanFic too, but I want Function calling and a RAG memory system first!
Damn, this is very cool.
like chatgpt talking mode but glados and integrated with smart home api! and running locally on two 4090s. awesome! now it just needs to pronounce GLaDOS correctly
Everyone gangsta until she floods the place with deadly neurotoxins...
oh wowww this us sooo cool
i love how humans always interrupt ais and expect them to just not care
Oh dear. As if this hobby doesn't slap you quite hard enough in the face already.
It's brilliant.
That is fuckin sick. Starred and waiting for long-term-memory and vision capabilities.
Excellent, now do it for SHODAN
Torment nexus type shit hahaha good work man!
Brilliant!!
Damn this is what Imagined people would make with language models, but the biggest problem is that most of answers are REALLY REALLY long.
Totally modifiable! I will make some vids onm tailoring the system to reply how to prefer!
This is amazing! It works flawlessly on my 1050 Ti with a 3B model. Will there be multilingual support, though? It seems that it's using Parakeet for transcribing speech, which supports English only. Maybe it's possible to use Whisper? I know it has been used before and I assume it was dropped due to its speed, but maybe it can be optional? The user might choose faster-whisper or similar for speed. And since you're planning to change the TTS model, maybe there's hope for a multilingual model too? Would be awesome, but it's fine if you decide to stick with English-only. Great project either way!
It's possible to switch, but my goal currently is the best possible system.
In a few months, theres sure to be better.and multilingual models. Once one pops up, I will swap it in!
@@dnhkng1 That's awesome! I'll be waiting, good luck on the project :)
@@dnhkng1 Kokoro is the latest hotnest in TTS but doesn't really support cloing much I think atm.
@@pylotlight Not cloning, but, you can finetune it. For a robot voice like GLaDOS, it would't be better (also according to the author of Kokoro).
I feel like GlaDOS would say no to your demands and would demand you to do things for her.
Coming, in the next update! She'll be able to start conversations with you!
This is awesome! Are you considering implementing this with a home assistant API?
I am working on improving the AI first. Once that's set up with function calling, if Home assistant can be accessed by API, it will work automatically.
GLaDOS is queen of condescention, you nailed it!
The computer name is “TwinTower” 💀
Neuro-sama at home lmao.
gee lad ee oh ess
so are we not gonna talk about the "TwinTower" hostname
Tower-Style PC Case, with twin RTX4090s. What's wrong with the name?
th-cam.com/users/shortsX31R_jO59L0
@@Only_Now_Matters It was a national tragedy
@@_zoid people today lack the moral guidance and upbringing required to care about other's lives. 🤷 I understand what you're saying though, it definitely was a tragedy, I wish the people lost and the families found peace.
@@Axodusit's been 23 YEARS GET OVER IT
Sounds good!
now i see why glados used neurotoxins
Will a russet potato be sufficient or do I really need two 4090s?
How did you get the tts to sound like glados with the distortion etc? One thing that bothers me with modern tts is that it always aims to sound human.
having GlaDos seemed impossible to me (have you tried what ever technique you used on a more challenging ai, SHODAN for example lol.
@@DaveTheAIMad fined tuned it myself! If I get enough interest, like a few K subscribers, I'll pause coding and make a TH-cam series on how to build GlaDOS, including training voices etc.
@@dnhkng1 well your 1 sub closer to that goal :). Its an impressive feat getting it to sound like GlaDos
The persona is just a tad too much Marvin from Hitchhikers Guide 😂
Yes, but the Lore is that she is stuck running on your Gaming PC GPU. Of course that's a bit annoying!
"Looks like you were trying to install the GNOME operating system, too bad"
GNOME is not an operating system, it's a desktop environment
@@lllIIIlIllIIllEh. Android and Chrome OS are colloquially considered to be their own operating systems, even though they both use the Linux kernel and Chrome OS uses a lot of GNU userspace. Heck, Steam OS might be considered an operating system, even though it's Arch Linux.
One might think binary compatibility defines what an OS is, except Windows has WSL1 where the kernel directly executes Linux syscalls at the same level as win32.
The lines are getting very blurry for colloquial usage. GNOME tries to define a large part of the user experience, with tons of inter-integrated components and recommended applications. For many users, desktop environment is a more important choice than OS distribution at this point.
@@antonliakhovitch8306 that's true, but everything you said about GNOME applies to any other major DE out there. And DE is definitely not the most important thing when starting out with Linux, try GNOME on Debian vs on Arch vs on Ubuntu vs on Mint, each will be a unique experience with vastly different sets of problems and advantages.
DE is merely a compatibility layer between the software and the user, one could say.
@@lllIIIlIllIIllthere is a GNOME operating system, which is mostly supposed to be an environment for testing and developing the desktop environment, but it exists
@@scarecat who in their right mind would install a test system?
Awesome work! I'm really enjoying the progress. If you have to use a model smaller, then llama3.3, which would you recommend
Just try a few, its a single line edit to the config to swap LLM models :)
What about cake? Is it lie or not?
Can a 2080TI on linux use it? Also more importantly can it do shit like making nextcloud work on docker and stuff like that?
nice! i wonder how the character of her voice can be improved, by having a bigger cohesion of spoken sentence and tonality/rhythm. I also think she yaps a bit too much. I wonder how 11Labs do it with their 11Listen app.
also is it using a wake word? or is it constantly listening in and able to reply once it detects you address it?
You can set up an optional wake-word. What cool is that wake word can be anywhere in the sentencem not just to 'wake' her up.
the only thing she lacks is the ability to say "no" to your requests
Thats amazing! Do i have to understand Python fully to build it for myself using ChatGPT (because im a broke student i only have a gtx 670) with your code?
Not at all! Just edit the glados-config.yaml file, and add the OpenAI address and API key.
Then add $5 to your OpenAI account, and she'll be up and running!
I'm not affiliated, and you can also use any OpenAI-compatible API provider with a llama model 🦙🦙 🦙
@@dnhkng1 Thanks for the fast answer. I will give it a try then :)
Haha this is fantastic!
But, what happens if you give her access to neurotoxins.
Already covered: github.com/dnhkng/GlaDOS/blob/9e3a2aa3f75e8253a4d474f996bb68f97b11f1ec/glados.py#L35
Great use of electricity lol
Are you also working on a Wheatly one?
Yes, that's in planning. Subscribe and you'll get it in your feed later.
Can you please collab with Neuro-sama
I need this
pls do the tutorial!!!!!!
Pro yapper.
There is no device for this program.
you might want to tell it to be "short and concise" so it doesn't make those very very long answers.. (edit: "Keep the responses as short as possible without breaking character.")
is there hal9000 onnx yet?
I might make one!
@@dnhkng1 cool. also what about the alice voice from resident evil
This is scary
Giladiaors
Damn i should make an nvim theme like this
Cool, finally, neurotoxin
I saw your personality core and I start thinking, it is possible to launch any usefull model on a smartphone? lets say for example Snapdragon 720G with 6GB of ram?
Should work with a tiny LLM, like Llama3.2-1B
With more RAM it's possible to run a 4B or even a 8B model.
I WANT THIS AI INSTEAD OF MY DUMB ALEXA!!