Block's Dessa AI Voice Is CRAZY: It LAUGHS Mid-Sentence (I Played the Real Demos)

Author: Prompt Engineer 48 · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Block’s Dessa AI Voice Is CRAZY: It LAUGHS Mid-Sentence (I Played the Real Demos) · Channel: Prompt Engineer 48 · 10:37 · Watch on YouTube

Topics: AI evaluation and benchmarks, Multimodal models, Privacy and data protection, AI safety

Summary

The video reviews Block’s Dessa speech-to-speech AI through recordings on its demo page, since the presenter cannot access it beyond a waitlist. The demos feature conversational overlap, laughter, expressive timing, and speech that changes pace with the mood, while Block claims the model was trained on human speech and can run locally. The presenter cautions that the recordings are curated, the claims lack public benchmarks or downloadable weights, and human-sounding voices could create safety risks. The verdict is that Dessa is a promising preview, not a product the presenter has independently tested.

Key points

Conventional voice assistants typically pass through speech-to-text, a language model, and text-to-speech, potentially losing vocal cues along the way.
Block describes Dessa as a speech-to-speech model that listens and speaks directly, with full-duplex conversation.
Block says the model was trained from scratch on human speech and can run faster than real time on a MacBook Pro M4 Max or better.
The demos include overlapping speech, laughter, brief acknowledgments, interruptions, and changes in pace and intonation.
The presenter says the timing in the anniversary surprise demo stands out, but cannot verify whether the emotions are genuine or performed.
The demo recordings are selected by Block, and the presenter notes there is no public paper, benchmark table, or downloadable model weights mentioned on the page.
A human-sounding voice could be used to pressure or imitate people, or to provide unhealthy emotional comfort; Block is seeking safety researchers to identify risks.
The presenter says local operation could enable voice applications without sending audio to a server, but also notes possible misuse such as scam calls.
Dessa is not available for general use in the video; interested users can join a waitlist.

Chapters

Start	Chapter	Summary
00:00	What Dessa claims to do	The presenter explains the limits of traditional voice assistants and introduces Block’s claims about Dessa’s speech-to-speech approach.
02:06	Listening to the demos	The presenter plays recordings featuring reactions, a surprise, everyday conversation, a movie-memory exchange, and life advice.
07:25	What stands out—and what remains uncertain	After briefly comparing a GPT live sample, the presenter recaps Dessa’s conversational qualities and raises concerns about curated demos and unverified claims.
08:32	Risks, potential uses, and verdict	The presenter discusses safety and privacy considerations, possible local uses, the waitlist, and why Dessa should be treated as a promising preview rather than a tested product.

Quotes

Interacting with current voice AI can just plain suck. — 00:31

It is a speech-to-speech conversational model. — 01:02

the timing is what stands out. — 04:42

a polished demo page is still a polished demo page. — 08:32

treat this as a promising preview, not a product review. — 10:06

Concepts and entities

Dessa, Block, Jack Dorsey, Grok, Raghavan Thirumalai Ratnam, RunPod, speech-to-speech, full-duplex

Full transcript (original language)

00:00 Imagine an AI that laughs in the middle of your sentence. Not because a script told it to, because you said something funny. That is the claim from a tiny lab inside Block, the company Jack Dorsey runs. And right now, it is trending on X. I could not try it myself because there is only a waitlist. So, in this video, I will do the next best thing. I will go through the whole page, play you the real demos, and tell you where I think the hype is real and where I would be careful. First, the pain. Talk to any voice assistant today and notice what actually

00:31 happens. Your voice goes in, a speech-to-text model turns it into words, a language model reads those words, and writes words back. Then a third model reads those words out loud. Three hops, and look at what falls on the floor on the way. How tired you sounded, the sarcasm, the little sigh before you said yes. All of that is gone before the smart part ever sees it. Block says it very bluntly on the page. Quote, “Interacting with current voice AI can just plain suck.” End quote. They say it throws away an incredible amount of information that text alone cannot carry. And the second pain is turn-taking. You talk, it waits. You stop, awkward gap. Then it talks, and

01:02 you cannot interrupt it. And it never says, “Uh-huh.” while you’re explaining something. So, what is Dessa? It is a speech-to-speech conversational model. Audio in, audio out, and no text in the middle. Here’s what Block says, and I want to be clear, these are their claims, not my measurements. It was trained from scratch only on human speech, no text, no pre-trained model underneath. It is full duplex, which just means it listens while it talks, like a phone call with a friend. So, it can say, “Yeah.” or “Got it.” or “Ooh.” while you’re still speaking. They say it hears and produces laughter, sarcasm, sighs, even vocal strain. It runs faster than real time, and it can run locally

01:35 on a MacBook Pro M4 Max or better. That last one matters. Local means your voice does not have to leave your laptop. I will come back to that. Now, why is it trending? X’s own summary, written by Grok, so take it with a pinch of salt, says the team calls it a breakthrough like GPT-3 for voice, and that some people find the voices a little too real for comfort. The team is small. The page calls it a very small AI research lab that reports directly to Jack Dorsey, founded by former members of Dessa, the Toronto deep learning startup. So, the name is an old company’s name. The lead is Raghavan Thirumalai Ratnam. Okay,

02:06 enough talking. The page has recordings of real people talking to Dessa. Each one has two tracks, one per speaker. I will play you short pieces, tell you what to listen for, and the credit stays on screen. First one, the release. A person says the team has decided to release the model, and the other person reacts. Listen to the reaction at the start. >> Oh, tell us. >> Yeah, we the team uh we’ve decided to release the speech-to-speech AI model we’ve built for the past >> Oh. Woah. I was So,

02:37 I have been watching your journey of creating this. >> And what what did you think of it? Are you excited about it as a listener? >> so blown away. I am so impressed, to be honest. >> right? >> It is so expressive. It it’s conversational and >> It’s It’s It’s very conversational. It can laugh. It can understand you and your emotions much better than anything I’ve seen out there. >> I haven’t really seen anything like that, either, but it’s >> Notice that it is not clean later. In that same clip, someone says, “You are a bit far from the mic. Can you come

03:09 closer?” That is how real calls sound, and Block left it in. >> Like uh I I would honestly ask it like >> you’re you’re a bit far from the mic. Can you come closer so I can hear you? >> I would be like, “What is consciousness? Is it something that’s programmed into us? Is it something that can be expanded on? Is it something that’s interdimensional?” >> That Those are all good questions, but I can Actually, I can’t really hear you. May Maybe you should cough up cough a little bit or something so you I could hear your voice better.

03:41 >> Sorry. >> Yeah, I think that’s better. Can you talk again? >> Uh >> Next, the anniversary surprise. Block’s own description is that Dessa gasps, “Oh my god.” under its breath, then sounds like teary gratitude. Listen for the gasp. >> That’s wonderful. So, I I have a surprise for you. >> Oh, thank god. >> Thank god. You you You were afraid I wouldn’t have a surprise for you? >> I was cuz I was like, I don’t know what to get you. And then I see all this TJ Maxx, and I saw this jumper and I’m

04:12 like, this is perfect for you to get. And then >> Okay, okay. But I have I have the best surprise for you for our our anniversary. >> [snorts] >> What? >> Okay, so you know how you’ve always talked about how much you want to go to Paris. >> Yes. >> So, for our anniversary I’m taking you on >> [clears throat] >> flight to Paris. >> I have to be honest, I cannot verify from a recording that every emotion is real versus very well performed. What I

04:42 can say is that the timing is what stands out. Timing is the thing the old three-hot pipeline cannot fake. Third, a plain everyday chat about coffee. No scripted wow moment. Listen to how one person tells a story and the other jumps in with a joke and an observation right on cue. >> Did you not Did you not learn from the first one? >> I thought [snorts] I did. But this one I was like, you know, let’s let’s be an extra safe cautious. >> Okay, I um I have an observation to make. >> Yes. >> I think that you’re a coffee addict. >> I I very much am.

05:16 >> [laughter] >> I also took two cups of black coffee this morning. >> Oh, yeah, yeah. How much coffee did you drink just today? >> I drank today about 15 cups. >> Okay, you’re going to die. >> I’m I’m well aware that I will probably die. This is me this morning. >> Now the movie memory one block says a person cannot recall a title and Dessa finishes the sentence. Listen for the moment where the two of them search for the name of the Batman movie together. >> I think Heath Ledger embodied the Joker.

05:52 >> Oh, uh this was in the second one, right? >> second one, yeah. >> Yeah. Uh which was that one called again? I I can’t >> The Dark Knight. >> Uh The Dark Knight >> Rises? >> No, Rises I I think is the third the one with Bane. >> Yeah. >> I I forgot what the second one was. Just The Dark Knight? >> I think so, yeah. >> Okay. But uh what other movies from Christopher Nolan do you like? >> The Memento. Memento is one of that I is a >> A Memento, you mean?

06:25 >> Yeah, Memento. >> Yeah, yeah, yeah. >> Life Advice, this one runs nearly 5 minutes on the page. I will play 30 seconds of it. It is slow and a bit heavy and Block says it sounds empathetic. >> experience and all. Like what what would have If if younger you was thinking the things I’m thinking right now, what would you say to her? I had I would just say to relax a little bit more. You know, I wasn’t um I didn’t always have straight A’s in

06:55 school, right? Like I wasn’t It wasn’t that serious and I I managed to get through, you know? I was I I’m pretty good at what I do and I don’t think I I would fail at anything right now, but it I I could have just eased up a little bit more and had a little not eased up a little just have a little bit slower pace of life. You know, that you know, back then it was always oh, I have >> Notice the speed changes. It slows down on the serious parts. Block calls at a wide range of intonation and speed.

07:25 There is also a GPT live sample on the page. Block notes that its developer is not affiliated with Block and has not endorsed Dessa and that the voices in those demos are AI generated. I cannot tell from audio alone which voice is which, so I will just play a short piece and make no claims. >> Let me check a few angles. There are a few big ways chat GPT style systems can get smarter. So, we’re going better training data, improved training methods like teaching reasoning strategies or better feedback loops, >> Mhm. >> and more compute. Like >> I mean, I’ve I’ve heard about the you

07:59 know, there’s uh Amazon GPUs and stuff out there now and all that stuff. I don’t know. I guess I’m just like kind of freaked out about what would what would be the way that chat GPT can learn? >> Quick recap of what you just heard. One, people talk over each other and the conversation survives. Two, laughter and small sounds like oh and yeah show up in the middle of sentences, not after them. Three, the pace changes with the mood. Those are exactly the things text-based voice assistants lose and that is why this page is turning heads. Now, let me

08:32 put my skeptic hat on because a polished demo page is still a polished demo page. One, these are Blocks chosen recordings. We do not know how many tries it took or what the bad ones sound like. The page says nobody steers it, which is a good sign, but it is still a curated set. Two, most of the big claims authentic, first, faster than real time are Blocks own words. The page itself says as far as we know there is no paper, no benchmark table and no weights to download. Three, emotion in audio is hard to judge by ear. A model can sound moved without anything inside. A nice voice does not mean it understood. The life advice clip is good listening, but I

09:05 would not take life advice from it. Four, and this is the one that actually worries me. Block itself is asking AI safety researchers to find risks it did not anticipate. A voice this human can be used to pressure people, to imitate them, or to comfort them in unhealthy ways. And one demo is two AI voices talking to each other. Imagine that on a phone line. So, why do I still think this is a big deal? Because of what it says about how to build voice AI. Blocks pitch is that one spoken word carries thousands of times more bytes than the same word written as text. All that extra is the tone, the pace, the breath.

09:35 Text throws it away. If a model learns from the sound directly, it can keep it. And they say it learned language from hundreds of times fewer words than text models need. If that holds up, it is a different road to talking machines, not text first and voice later, voice first. It is also practical if it really runs on a laptop, a voice agent that never sends your voice to a server becomes possible language practice, tutoring, accessibility, and yes, also scam calls. Can you use it today? No. There is a join the waitlist button on the page and a separate form for safety researchers which needs an institutional email. Blog

10:06 says more recordings are coming on the X account of the lead, Raghavan Thurairatnam. There is also a privacy notice effective October 21st, 2026. I would read it before joining anything that records your voice. It says they collect identification and device data, do not sell your information, and may store data in the US, Canada, Japan, and the EU. My verdict, what I heard, natural overlap, interruptions, laughter, and speech changes convincing in a way the old pipeline never is. What I did not get, any test I could run myself. So, treat this as a promising preview, not a product review. If it

10:37 ships and weights or an API show up, I will test it properly. Interruptions, background noise, accents, long conversations, and I will show you the failures, too. If you’re looking for cheap and reliable GPUs, I use RunPod, and the link is in the description. Tell me in the comments, would you talk to an AI that sounds this human, or does that creep you out? Like the video if it helped, and I will see you in the next one.