alt Hacker NewsDessa: Conversation Model trained only on speech
Comments
These bidi/duplex demos always come off as overly artificial to me. It's cool that the model can speak at the same time as you instead of taking turns, but training it to interrupt as much as possible isn't the way to do it. In the very first example, did anyone here find the laugh after the "winter" answer to be convincing? What was the need for it? Why did the human speaker not get to finish a single sentence or finish a thought before the model took over with a monologue?
These full-duplex models are getting very "Her"-ish lately. gpt-live-1 is fantastic. Dessa, though, seems to be aiming at full human parity. There's a huge market for AI girlfriends that they should be looking into in earnest...
Most recent voice-related tech is using voice as an interface or interaction mode. It’s a means to an and, and the end is the value provided by the reasoning system on the other side of the voice.
This model is clearly not so smart. So the voice is the whole point, there’s not much value beyond it, at least on its own.
What that adds up to is, impersonation. That’s what this is for, impersonating humans.
It sounds like it is trained on a lot of podcasts, judging by a quick few second skims over their demo tracks.
Also, the trend of posting on HN with much more sensationalist title first, getting to the front page, then getting the title adjusted by a moderator should probably be penalized; how is 'most authentic ever' not a marketing term? (not just this post, but a bunch of posts recently got to frontpage by making unreasonable claims)
I thought it sounded a little stilted and weird and then I realized that was the human half of the conversation.
The conversation with the AI fake wife was a little creepy, albeit honest about what this tech is probably going to be used for.
There's no way I would be able to detect if this was AI from the examples alone. If I got a phone call from a malicious agent using this tech, I probably wouldn't realize it. My plan is to ask a challenge question about a shared memory with people I personally know, if there's something suspicious about a call. I don't know how else to prevent it. It sucks to be this suspicious, but I can't help it.
There's so many potential scenarios. One scenario: An AI voice trainer trains on someone's social media video posts and get a pretty perfect copy. Then SIM jacks the target's phone number and calls pretending to need help. AIs could be used to mock ransom people, request bail money, ask for a change on a production server, etc.
People with high public visibility are likely targets, especially if they have money and or power. Impersonating them to subordinates would be an easy means of a social engineering attack.
The voice seems insanely immersive and uncanny. One issue that the full duplex models always have is tool calling. This is where GPT-Live [1] shines in the sense that you can use it as a fully practical agent. Looking forward to seeing how Dessa folks implement this.
Very technically impressive, but the fake conversation with the machine was cringe worthy to me; this pretending to care doesn't make me want to converse at all. The goal for me is to have an intelligent conversation that _adds_ to my experience; whatever I'm working on at the time. The AI girlfriend experience makes me cry inside rather than rejoice.
> LLMs need an insane amount of text. Dessa learned language from hundreds of times fewer words (in fairness, LLMs also have to learn C++ and whatever the hell Angular is).
Now that a larger proportion of text is AI, I'm appreciating human-written text more and more; this got a good chuckle out of me.
Sounds like they trained it on Boxxy or Kat from Sam and Kat. I guess it does sound like a human talking but in an affected unnatural way.
Wondering what kind of tokenization they use after the model referred to the movie Memento as The Mento the first time before being corrected.
I was expecting to run into some uncanny valley-esque. It was surprisingly good.
It is kinda scary that we're here already, I can see this being pretty problematic for seniors/public prone to getting scammed.
> we figured out how to train Dessa on only one NVIDIA GB200 NVL72.
This is the crazy part. It seems pretty smart, and LLMs require 1000x more compute to train. Can this kind of efficiency somehow be achieved with text too?
Passing a Turing test over a 5 minute conversation:
- Text interface: Done
- Voice interface: 5-10 years I’ll predict
The demo is nice, but enough obstacles remain that very few people would mistake it for a human over even a 5 min conversation. Chat systems configured to intentionally impersonate can already get people to mistake them for humans most of the time.
There’s been great progress in this area, direct response modeling, parallel reasoning, but last challenges are huge.
The reality is LLMs still seem dumb much more often without significant reasoning time (test time inference). On that point alone they are still an order of magnitude too slow for convincing interactive latency.
Of course they don’t need to meet this bar to be useful, systems are good enough today for lots of applications. It will just be interesting to see what the final delta ends up being.
So, I tolerate the "ums" and "ahs" and "uhs" in conversation from humans because they're human. If an AI is going to speak to me (gods forbid), it better be concise and clear and focused, or I'm throwing it out the fucking window.
It is interesting that they've done this reasonably successfully, impressive technology. Novel ideas can have inherent value, absent direct uses they may have, as they can be a stepping stone toward other inventions. But, I'm having a hard time coming up with a use case that isn't predatory or intentionally misleading. Company support agents pretending to be human, for instance. Or, taking the next step on AI girlfriend/boyfriend type interactions. Or, outright scams. Models already do an excellent job with voice cloning, at least for intonation and timbre, adding realistic conversational tics makes it even more likely grandma is going to fall for it when her "nephew" calls needing help.
But, that seems to be extremely common with new tech these days. There's always an angle that is unsavory. Marketing, scams, propaganda, surveillance, killing, that's where the money is, so everything bends toward them.
Close enough, welcome back Google Duplex!
https://www.youtube.com/watch?v=D5VN56jQMWM
For real though, it's been bothering me, how did they do this in 2018? (Some kind of symbolic logic based AI?)
Also props to the people at Block, Dessa is very cool :)
I know someone who recently came across a scammer where a guy impersonated a sheriff's deputy and tried to get money out of her. She told me the guy sounded like a real sheriff and made her believe the guy until he asked her to pay some money in Bitcoin. I can imagine that this type of scam could go viral soon. Even worse, someone could record our voice and fine-tune a model like this so that it can produce our voice and then talk to our family members to deceive them. It'd be a real trouble.
This feels like a parlor trick where the AI was trained to mimic informal telephone mannerisms. The actual conversations were pretty bad. Lots of style, not much substance basically.
The real question is, does it actually understand sound?
Conventional LLMs are just running speech to text and then working on the text, so they can’t understand music or humming or any sort of sound.
Open weights or it didn't happen.
Really impressive demo. I have to say that, but I do worry with this kind of fidelity to a human interaction that people will get more drawn into AI relationships and replace human contact.
I like voice as a mode of interaction with AI, but it’s always been clear with current tech that I’m talking to AI (and yet folks still got drawn into fantasy relationships thanks to sycophantic models etc) - I imagine with voice models like Dessa that could be worse for humans. This kind of fidelity blurs the lines of who or what the human is talking to and lets humans fall into an isolated world where they believe they are “socializing.”
Now, we could argue this is not a negative, maybe it is actually healthy for those who are looking for social connection or lonely, but my gut feel is that is not the case for most. It worries me that this kind of model will actually cause further deterioration of human socialisation and human-to-human contact.
If you actually want your "AI" to interact with anything digital, are these voice-to-voice models going to...use speech to text? How does this model interact with anything that isn't an audio channel?
I actually thought it sounds really good and natural and it freaked me out in a good way. You can tell that the robots have no idea of what they're talking about, but the voicing, the emotions, etc, really good.. Like talking to a friend on the phone. Also creepy, because of the potential for abuse of this technology.
But the interesting part to me is that training on voice only is what allows models to learn emotions - voice transmits a lot of emotional information which text doesn't capture. Language has words for a very limited number of emotions - maybe only the peaks - but there's so many more which we can't describe or transmit via text.
You could determine the phonetic tokens which express different kinds of emotions or 'render' text+emotion using a DSL, etc. Exciting stuff.
Wanted to joke "AI girlfriend just dropped" but they explicitly have an AI girlfriend as the 3rd example.
Modern problems...
it only works for english now, right? right?
It sounds good, but...what the hell is it talking about?
The "Huge range in intonation and speed." example. "I had another ollie special". "I ordered a French vanilla latte, special from Maude [or is that Aldi?], springs and waffles, granny smith"...wtf does any of this mean?
She takes a sip of something and her "mouth is on fire...from all the sugar". She repeatedly calls herself "a ding".
Third example at 0:02:
>Dessa: Hey, baby.
Block is in the AI girlfriend market? Ew.
Interesting tech, but if these are the hand-picked examples that best represent it, I suppose we're not quite there yet.
The intelligence sounds lacking. At this point, anyway.
But the idea of learning through voice instead of written text might prove fruitful. After all, text isn't strictly native to humans. It's just a tool we've invented.
Why is every comment a brief "very impressive, nice job" on a high impact technology followed by a tangential semi-personal rant?
Not bad, but I guess I lack the imagination to envision the use cases for such casual conversation. Other than... I suppose a certain demographic would be lining up to use this as the voice for their AI girlfriend.
Also I seem to be missing where/how they are selling this.
Both speakers in the examples sound unnatural. The human doesn't know what to talk about and seems to have little confidence. The robot says things that only mostly make sense and draws on, saying ultimately very little. "How's the winter stuff?" e.g. is not a sentence a person would say. I might say "How are you dealing with the snow?" or "Are you looking forward to winter?," but not "How's the winter stuff?" And there's just so many examples like this. Also, I'm not convinced that this "human-like" conversation with filler words and interruptions and pauses is actually what anyone wants, especially in absence of a face to interpret or a real human being to understand behind it. Even if it worked perfectly, it would feel uncanny and artificial because I know the person isn't real.
Can you instruction-train a duplex model like this? How would that look? I also wonder, if a model trained only on speech can develop world knowledge like this, could you train an image model only on images of text to essentially act as an LLM?
ya if this is real this is magnitudes better than anything else out there, it honestly feels like it can't be real because its too real, i would be 100% fooled and never even question it wasn't a real person, i was more worried because of that, in that no one would be able to tell and you could have very engaging / emotional conversations that chatgpt/grok/claude etc can't even come close to
This is so awesome! How does it work with moving away from the mic, is it just random or does the conversational pattern lead it that way?
The volume on this seemed so low and wouldn't get better even on max on my MacBook Air M5. Hoping it isn't my ears lol
Ok. Sounds cool. Now what? Open source weights? An API somewhere? Available on Openrouter? Are they just testing?
Where is this going?
reminds me of the last year hype around SesameAILabs/csm, people losing their mind over how realistic it was, only for them to disappear and never ship. sorry can't get hyped over some mp3s.
Great. I wonder.. is this based on Mochi or PersonaPlex? Also, how big is the model, and do we use it?
not at all creepy
But can it follow a simple customer support script?
Now that a robot can speak over people on Zoom calls for me, I’m finally free to write poems all day!
Call me a skeptic, obviously the voice sounds plausible and very realistic and human-level but where does the intelligence come from if it's only trained on human speech datasets? LLMs are powerful because they can do anything, if you only have a voice corpus I doubt it is as powerful as all written human knowledge. Like when it says "My day was good", what does it actually mean? Or is it just generating plausible slop.
Very impressive demo!
It is cool but since there is no way to confirm ie. run it, or see it working somewhere, it can be also not real.
I get you are excited to share, but taking into account what your product is, it would be more powerful to let people experience it. As a creator of digital products, why would you put so much text on the homepage, anyhow you get the point.
The speech is great, but you can see some cracks even in the first sound bit. near the end he asks the model to speak louder and initially the voice is louder but then goes back to the same level, also losing what she was saying just seconds ago and he has to lead back to the topic.
I'm surprised it took this long! Clearly the next step is joint training on text and audio. The speech realism is very good, but the intelligence is pretty clearly 2024-level.
The conversation examples are impressive but this made me say holy shit out loud: "it runs locally on a MacBook Pro M4 Max or better. And we figured out how to train Dessa on only one NVIDIA GB200 NVL72."
If it really can replicate that low latency + that expressiveness on local devices, that's one heck of an accomplishment for "indie" AI.