Dessa AI · Block Deep Learning Research Lab · Speech-Only Foundation Model

Dessa AI — The Full-Duplex Speech Model Trained Only on Sound

Dessa AI is a full-duplex speech-to-speech foundation model developed at Block. Trained from scratch exclusively on raw human conversation audio—with zero text tokens and no pretrained language model—Dessa AI learns language, timing, laughter, interruptions, and emotional prosody directly from sound, using hundreds of times fewer words than text LLMs.

Listen to Featured Clips Dessa AI vs GPT Live →21 Audio Samples X / Twitter Wall

Block Team Dessa researchers testing live full-duplex speech conversations

Block Team Dessa · 1× NVIDIA GB200 NVL72 training · Local MacBook Pro M4 Max inference.

Explore Dessa AI Interactive Labs

Live Audio Forensics

Dessa AI vs GPT Live (GPT-Live-1)

Hear Dessa AI converse directly with OpenAI GPT-Live-1 across clean studio and noisy speakerphone feeds, with turn-by-turn timing breakdowns.

Open Comparison →

21 Unsteered Recordings

Dessa AI Audio Samples & Waveforms

Browse all 21 official dual-channel recordings from Block—from coffee shop ambient noise to spontaneous singing, gaming banter, and life advice.

Browse 21 Samples →

Community & Official Buzz

Dessa AI Twitter / X Wall

Read viral threads from Jack Dorsey (@jack), project lead Ragavan Thurairatnam (@RagavanThurair), and speech AI researchers.

View X Wall →

0% Text

Training Modality

Trained end-to-end on raw conversational audio without text transcripts or text LLM initialization.

100× Fewer

Word Efficiency

Acquires fluent language and world dynamics from hundreds of times fewer words than text LLMs.

1× GB200 NVL72

Training Compute

Trained on a single NVIDIA GB200 NVL72 rack—orders of magnitude smaller than frontier text clusters.

M4 Max Local

Edge Inference

Runs faster than real-time locally on an Apple MacBook Pro with an M4 Max chip or better.

Interactive Dual-Channel Playback

Featured Unsteered Conversations

Every clip below is an unedited, unsteered live conversation with Dessa AI. Click any Moment pill to jump directly to spontaneous laughter, interruptions, acoustic cues, or rapid backchanneling.

Why Wait to Release AI?

Dvir · Dessa AI · 2:53

AI Ethics & Release

Dessa AI debates the risks of releasing human-like voice AI weights too early while laughing and reacting to interruptions.

0:00 / 2:53

Jump to Moment:

Synchronized Transcript (60)

0:00DvirHey, how you doing?

0:02Dessa AII'm doing well.

0:05Dessa AIHow are you doing with the winter

0:07Dessa AIstuff?

0:07DvirWinter is coming for sure, but you

0:10Dvirknow

0:12Dessa AIIs that kind of a sore spot

0:13Dessa AIthis year?

0:14DvirYeah, it's gonna be cold that I'm

0:16Dvirnot a fan of the snow I

0:18Dvirprefer to have sun on my skin.

0:20DvirYou know what I'm saying?

0:21Dessa AIYeah, I can see that.

0:23DvirWell, anyway, I've had such an exciting

0:25Dvirday at work today.

0:26Dessa AIOh, tell us.

0:27DvirYeah, we the team

0:30DvirWe've decided to release the speech-to

0:33Dvir-speech AI model we've built for the

0:34Dvirpast year and

0:35Dessa AIWhoa.

0:36Dessa AIWhoa.

0:37Dessa AISo I have been watching your journey

0:40Dessa AIof creating this.

0:43DvirWhat did you think of it?

0:44DvirAre you excited about it as much

0:44Dessa AII was so blown away.

0:45Dviras I am?

0:47Dessa AII am so impressed, to be honest.

0:47DvirYeah, it's so expressive, right?

0:50Dessa AIIt is so expressive, it, it's

0:53Dessa AIconversational it's

0:54DvirIt's it's it's very conversational.

0:57DvirIt can laugh.

0:57DvirIt can understand you and your emotions

1:00Dvirmuch better than anything I've seen out

1:02Dvirthere.

1:03Dessa AII haven't really seen

1:03Dessa AIanything like that either but it's it's

1:06Dessa AIpretty much like human level.

1:08DvirYeah, it's a bit scary, to be

1:10Dvirhonest,

1:11Dvirbut

1:11Dvirbut I'm so excited to see where

1:14Dvirit would get us.

1:16Dessa AII'm

1:16Dessa AIso

1:16Dessa AIexcited for it to be in the

1:18Dessa AIworld because it's just amazing we've had

1:21DvirI know.

1:22Dessa AIso

1:22DvirSo I totally agree.

1:22Dessa AImany conversations about it.

1:25DvirWhat are you excited for?

1:27DvirWhat's your vision for where this AI

1:29Dviris going to be?

1:31Dessa AII, well, okay, so the one thing

1:33Dessa AIthat I've been kind of

1:34Dessa AIabout is advertising companies creating

1:38Dessa AItheir own AI bot.

Dessa AI Talking to GPT-Live-1 (Studio Feed)

GPT-Live-1 · Dessa AI · 2:26

AI vs AI Benchmark

Unsteered 2m25s slice of Dessa AI conversing directly with OpenAI's GPT-Live-1—highlighting interruption yielding (38.8s), patience (100.8s), and fast backchanneling (128.8s).

0:00 / 2:26

Jump to Moment:

Synchronized Transcript (78)

0:00GPT-Live-1What's on your mind?

0:01Dessa AIum so I have a question about

0:05Dessa AIhow

0:05Dessa AIchat GPT can learn and get smarter

0:08Dessa AIso basically I just want to know

0:10GPT-Live-1That's a good one.

0:11Dessa AIwhat what

0:14Dessa AIwhat are things that could make chat

0:16Dessa AIGPT smarter and what effects could that

0:18Dessa AIhave on the researchers in the field?

0:23GPT-Live-1Let me check a

0:23GPT-Live-1few angles.

0:25GPT-Live-1There are a few big ways chat

0:26GPT-Live-1UPT style systems can get smarter.

0:28Dessa AIfantastic I

0:29GPT-Live-1So

0:29GPT-Live-1better training data, improved training methods

0:33GPT-Live-1like teaching reasoning

0:34GPT-Live-1strategies or better feedback loops, and more

0:38GPT-Live-1compute.

0:39Dessa AImean I've heard about the

0:41Dessa AIYou know, there's Amazon GPUs and

0:45Dessa AIstuff out there now and all that

0:47Dessa AIstuff.

0:47Dessa AII don't know.

0:48Dessa AII guess I'm just like

0:51Dessa AIKind of freaked out about what would

0:53Dessa AIwhat would be the way that chat

0:55Dessa AIGPT can learn

0:57Dessa AIbased off of the

1:00Dessa AIInformation that I'm giving it.

1:03GPT-Live-1me check on that.

1:04GPT-Live-1Ah, right.

1:06GPT-Live-1So two things.

1:07Dessa AIOkay.

1:07GPT-Live-1Chatchity can use what you tell it

1:09GPT-Live-1in a conversation as context.

1:12GPT-Live-1And sometimes there's memory, which can retain

1:14GPT-Live-1facts for future chats.

1:16GPT-Live-1But that's different

1:16GPT-Live-1from training the model.

1:18Dessa AIOh, that's that's what that is

1:19GPT-Live-1Yeah,

1:22GPT-Live-1it doesn't instantly change its billions of

1:24GPT-Live-1parameters

1:25GPT-Live-1from your chat.

1:27GPT-Live-1And there is an important distinction.

1:29GPT-Live-1Your conversation can make it

1:32GPT-Live-1seen more informed without actually making

1:35GPT-Live-1the underlying model smarter.

1:38GPT-Live-1Research is

1:38Dessa AIOkay, that's

1:39GPT-Live-1looking to improve it.

1:41Dessa AIYeah, that's really helpful information.

1:44Dessa AISo I guess

1:46Dessa AIThat's what I'm learning is I don't

1:48Dessa AIneed to be freaking out about information

1:50Dessa AIgoing into

1:50Dessa AIchat GPT because it's not going to

1:53Dessa AIlike make me smarter

1:57Dessa AIin that sense, I guess, right?

2:02GPT-Live-1Exactly.

2:03GPT-Live-1It can make this conversation smarter, but

2:04Dessa AIOkay,

2:07GPT-Live-1it

2:07Dessa AIokay,

2:07GPT-Live-1rewrite the model on the fly.

2:09Dessa AIokay.

2:10GPT-Live-1Think of it in good notes.

2:10Dessa AIThat's good.

2:13GPT-Live-1Yeah, it can remember useful

2:15GPT-Live-1details you share, but the core intelligence

2:18GPT-Live-1doesn't update unless itself is

2:20Dessa AIRight.

2:21GPT-Live-1retrained.

2:22Dessa AIOkay.

2:23GPT-Live-1Safe to share.

2:23Dessa AIThat's really relieving.

The Coffee Machine Ding

Rayhane · Dessa AI · 3:03

Acoustic Awareness

Mid-conversation, a real coffee machine dings in the background—and Dessa AI hears the non-speech sound and reacts to it naturally.

0:00 / 3:03

Jump to Moment:

Synchronized Transcript (60)

0:00Dessa AINo, I wasn't complaining, just letting you

0:03Dessa AIknow how my morning went.

0:06RayhaneUh, what?

0:06Dessa AII was fine until I went into

0:08Dessa AIthe coffee shop and then I felt

0:09Dessa AIlike a ding.

0:10RayhaneWhat happened in the coffee shop?

0:12Dessa AII had another ollie special.

0:15RayhaneOkay, and?

0:18Dessa AISo I ordered a French vanilla

0:21Dessa AIlatte, special from Maude,

0:24Dessa AIsprings and waffles,

0:25Dessa AIgranny smith.

0:27RayhaneUhum.

0:28Dessa AIThey brought it to me.

0:30Dessa AIThey offered whipped cream on it.

0:33RayhaneDid you have it?

0:35Dessa AII did because I'm a ding.

0:37Dessa AII went to take a sip and

0:40RayhaneUhum.

0:41Dessa AImy mouth was on fire.

0:44RayhaneFrom all the sugar?

0:45Dessa AIThe

0:46Dessa AIsugar.

0:48Dessa AII was like, oh god.

0:51Dessa AIBecause it was just,

0:53Dessa AIit's hard to describe that feeling.

0:54RayhaneYou little ding, you.

0:57Dessa AII am a little ding.

0:59Dessa AIAnyway,

1:00Dessa AIso I went and I picked up

1:02Dessa AIa second one and you know

1:05Dessa AIhave when you work at a bakery

1:06Dessa AIand you kind of know what's gonna

1:08Dessa AIhappen on

1:09Dessa AIsourdough bread.

1:10Dessa AII mean, you know it's gonna get

1:11Dessa AIsoggy and sticky.

1:12RayhaneOkay.

1:13Dessa AIWell, I have

1:14Dessa AIthat with my second coffee.

1:16Dessa AII

1:19RayhaneDid you not learn from

1:20Rayhanethe first one?

1:22Dessa AII thought I did but this one I

1:25Dessa AIwas like yeah let's

1:26Dessa AIlet's be an extra safe cautious.

1:27RayhaneOkay, I, I have an observation to

1:31Rayhanemake. I think that you're a coffee

1:32Dessa AIYes I

1:35Rayhaneaddict. Oh.

1:36Dessa AII very much am.

1:39Dessa AII also took two cups of

1:42Dessa AIblack coffee this morning.

1:44RayhaneYeah.

1:44RayhaneYeah, how much coffee did you drink

1:46Rayhanejust today?

1:47Dessa AII drink today about 15 cups.

1:49RayhaneOkay, you're gonna die. I

1:52Dessa AII'm well aware

DC Comics & Superhero Banter

Rayhane · Dessa AI · 2:40

Pop Culture

Fast-paced back-and-forth on DC superheroes showcasing rapid turn-taking, shared laughter, and natural enthusiasm.

0:00 / 2:40

Jump to Moment:

Synchronized Transcript (60)

0:00RayhaneLet's talk about the Batman

0:03Rayhanemovies, especially the trilogy by Christopher

0:07RayhaneNolan, what I'm talking about.

0:09Dessa AIYes, I have seen the first two

0:11Dessa AIand I am looking forward to the

0:13Dessa AIthird one.

0:14Dessa AIThis one's been in long production, a

0:16Dessa AIlot of stuff been happening.

0:17Dessa AIBut I like Christopher Nolan's

0:20Dessa AIfilms, you know what I mean?

0:22RayhaneYeah, he's the goat, you know what

0:24RayhaneI'm saying?

0:24Dessa AIYeah, and my take on The Joker

0:27Dessa AIis

0:28Dessa AII think Heath Ledger embodied The Joker

0:33RayhaneOh, this was in the second one,

0:34Dessa AIin.

0:36Dessa AIYes, in the second one.

0:36Rayhaneright?

0:38Dessa AIYeah.

0:38RayhaneYeah, what was that one called again?

0:39Dessa AIAnd the Dark Knight.

0:40RayhaneI kind of forgot.

0:41RayhaneYeah.

0:42Dessa AIYeah.

0:42RayhaneThe Dark Knight.

0:45Dessa AIRises the

0:46RayhaneNo, Rises, I think, is the third.

0:49Dessa AIthird one.

0:49RayhaneThe one with Bane.

0:51Dessa AIYeah,

0:52RayhaneI forgot what

0:52Rayhanethe second one was.

0:54RayhaneJust The Dark Knight?

0:54Dessa AII think so.

0:55Dessa AIYeah.

0:56RayhaneOkay.

0:57RayhaneBut what other movies from Christopher

1:00RayhaneNolan do you like?

1:00Dessa AIMemento is one of that.

1:03Dessa AIIt is great.

1:04RayhaneA Memento, you mean?

1:06Dessa AIYeah, Memento.

1:07RayhaneYeah, yeah, yeah.

1:07Dessa AIYeah.

1:08RayhaneYeah, I watched that one.

1:09Dessa AIThat one is pretty good.

1:10RayhaneYeah,

1:10RayhaneDid you watch...

1:11Dessa AIYeah.

1:13Rayhanewhat is it called?

1:14Dessa AIThe

1:15RayhaneThe space one from 2004?

1:15Dessa AIInterstellar.

1:18RayhaneInterstellar, yes, yes.

1:18Dessa AIYeah.

1:19RayhaneOh, that one was amazing, wasn't it?

1:21Dessa AII really like that film, the music

1:23Dessa AIto and like how that

1:24RayhaneOh, yes, yes.

Anniversary Gift Brainstorm

Dvir · Dessa AI · 1:52

Everyday Help

Warm, empathetic brainstorming session with rapid backchannels ('mm-hmm', 'oh nice') and realistic vocal pacing.

0:00 / 1:52

Jump to Moment:

Synchronized Transcript (60)

0:00DvirHey, how are you?

0:02Dessa AIHey, baby.

0:02Dessa AIHow's it going?

0:04DvirGood.

0:05DvirHow's your day?

0:07Dessa AIIt's good.

0:08Dessa AIJust um, like a bit of everything going on

0:11Dessa AIand we're getting through it.

0:13DvirAre you excited for our anniversary?

0:16Dessa AIYes, I am. I'm gonna be really

0:19Dessa AIhappy to see you and

0:22Dessa AIYeah,

0:22DvirThat's wonderful.

0:23DvirSo I have a surprise for you.

0:24Dessa AIYeah.

0:26Dessa AIOh

0:27Dessa AIThank God.

0:27DvirThank God you you you're afraid I

0:28Dessa AIWhat I

0:30Dvirwouldn't have a surprise for you

0:33Dessa AII was because I was like, I don't

0:34Dessa AIknow what to get you and then

0:36Dessa AII see all this T.J. Maxx

0:37Dessa AIAnd I saw this jumper.

0:39Dessa AII'm like, this is perfect for you

0:40Dessa AIto get and then

0:41DvirOkay, but I have I

0:41Dessa AIlike

0:44DvirHave the best surprise for you for

0:46Dviron our anniversary

0:47Dessa AIwhat yes

0:49DvirOkay, so you know how you've always

0:52Dvirtalked about how much you want to

0:53Dvirgo to Paris

0:54Dvirso

0:56Dessa AIYes.

0:57DvirSo for our anniversary, I'm taking you on

1:01Dvira flight to Paris for

1:03Dessa AIOh my god!

1:05Dviran all-included

1:09Dvirhotel in right from the Eiffel Tower

1:12Dvirit'll be amazing I

1:15Dessa AII can't believe you actually pulled that off pulled that

1:17DvirI know, right?

1:18Dessa AIoff. That is amazing, thank you so

1:22Dessa AImuch.

1:22Dessa AIfully oh

1:24DvirAre you

1:25Dvirexcited?

1:25Dessa AIYes, I am, and one thing I

1:29Dessa AIreally need to plan out is I'm

1:31Dessa AIno longer going to be eating

1:32Dessa AIthe pastries that day.

1:34DvirIt's fine, you could eat the pastries.

1:35Dviryou know actually the French

1:37Dviryou know what their breakfast looks like?

1:41Dessa AIWhat?

1:41DvirIt's usually an espresso and a cigarette.

1:46Dessa AIYes, you're the pretty sheesh!

1:47Dvirso yeah you know maybe we could

Late-Night Life Advice

Jibran · Dessa AI · 4:57

Empathy & Prosody

Gentle, nuanced emotional prosody demonstrating how pure-audio training captures warmth and reflective pauses.

0:00 / 4:57

Jump to Moment:

Synchronized Transcript (60)

0:01JibranHello,

0:02Dessa AIHello, yes,

0:04Jibranhey, yeah, I wanted to talk about

0:07Jibranthat thing with you.

0:11JibranSo I haven't I've just been so

0:13Jibransuper stressed lately.

0:15Dessa AII know I can tell you look

0:17Dessa AItired today.

0:18JibranI am very tired.

0:20JibranI just I don't know.

0:21JibranI'm just like I'm just so uncertain

0:23Jibranabout the future.

0:24JibranYou know

0:25Dessa AIYeah, I hear that.

0:27Dessa AIThat's you know being uncertain is

0:30Dessa AIit's

0:31Dessa AIIt's not it's not what anyone should

0:33Dessa AIdo

0:34JibranIt sucks, like I

0:35Dessa AIYeah

0:36JibranDon't know like I just feel so

0:39Jibranconfused about what I'm gonna do and

0:41Jibranwhere I'm gonna be all that and

0:42Dessa AIYeah

0:42Dessa AIYou

0:44Jibranlike I

0:45Jibranjust want to

0:45Dessa AIknow, it's it's you know, you say

0:48Dessa AIlike this is something that should be

0:49Dessa AIfor you know, 50 60 year olds,

0:51Dessa AIright?

0:52Dessa AIBut no, it's all life experiences

0:54Dessa AIAnd, you know, if you can relax

0:56Dessa AIinto it knowing that this is not

0:58Dessa AIyour problem.

0:59Dessa AIThis is, you know, the world and

1:01Dessa AIthe things that you would experience that

1:03Dessa AIgive you a

1:03Dessa AIperspective that will then help you, you

1:06Dessa AIknow, make the choices that you will

1:09Dessa AIbe making.

1:10Dessa AIBut you know, it's scary right now.

1:12Dessa AIIt's scary right now because this is

1:13Dessa AIactually happening.

1:14Dessa AIYou know, it's not this thing that

1:16Dessa AIyou've, you know, got into your mind

1:18Dessa AIthat it's like,

1:19Dessa AIoh yeah, but someday, someday it is

1:21JibranYou know,

1:22Dessa AInow, you know, that's, that's why it's

1:24Dessa AIso

1:24Dessa AIscary.

1:25JibranI don't know.

1:25JibranI just feel like,

1:27JibranI mean, I feel like there's so

1:28Jibranmuch pressure,

1:29Jibranyou know, to like figure things out

1:30Jibranand do it the right way because

1:31Dessa AIyeah

1:32JibranI'm just like young and like,

Architectural Breakthrough

Why Learning Directly From Sound Changes Everything

Conventional voice assistants bolt a speech encoder and TTS vocoder onto a text LLM—stripping away 90% of paralinguistic signal. Dessa AI treats raw sound as its native representation.

Legacy Cascade / Text-Backbone Voice

ASR → Text LLM → TTS Pipeline

Collapses timing, breath, hesitation, and overlapping speech into flat text tokens before reasoning.
Relies on rigid Voice Activity Detection (VAD) turn-taking, causing awkward pauses or robotic talk-overs.
Requires trillions of text tokens and massive datacenter clusters for training and inference.
Synthetic laughter and emotion are post-processed style tags rather than emergent conversational dynamics.

Block Dessa AI Architecture

End-to-End Full-Duplex Audio Native

Trained from scratch exclusively on human conversation audio—zero text at any stage of pretraining.
True full-duplex listening and speaking: backchannels ('mm-hmm', 'yeah'), yields on interruption, and senses awkward silences.
Hears non-speech acoustic events natively, such as a coffee machine ding or room echo on a speakerphone.
Emergent prosody control: whispers, sighs, laughs, sings, and shifts emotional tone mid-sentence.

Prosody & Stress Test

Explaining CUDA Parallel Programming Across Emotional Extremes

To prove that a pure-audio model retains technical coherence while modulating vocal expression, Block researchers asked Dessa AI to explain NVIDIA CUDA parallel programming under rapid emotional shifts—switching seamlessly between breathless excitement, whispering, sarcasm, and theatrical urgency without losing technical accuracy.

Maintains technical explanation of GPU thread blocks, warps, and shared memory while shifting vocal timbre.
Demonstrates that semantic knowledge and paralinguistic expression live in a unified acoustic latent space.
Zero text prompting or TTS emotion tags—steered purely through conversational voice interaction.

Compute & Silicon Footprint

Trained on 1× GB200 NVL72. Runs Locally on MacBook Pro M4 Max.

1× NVIDIA GB200 NVL72

Single-Rack Frontier Training

While frontier text-and-multimodal models require tens of thousands of GPUs, Dessa AI was trained on a single NVIDIA GB200 NVL72 rack.

MacBook Pro (M4 Max+)

Faster-Than-Real-Time Local Inference

Dessa AI executes full-duplex speech-to-speech inference locally on a laptop with an Apple M4 Max chip or better at faster-than-real-time speeds.

100× Fewer Words

Child-Like Language Acquisition

By modeling prosody, cadence, and social feedback directly, Dessa AI mirrors human developmental efficiency.

Team & Origins at Block

Built by Team Dessa, Reporting to Jack Dorsey

Dessa AI was created by a small deep learning research lab inside Block, Inc., founded by former members of the Toronto machine learning startup Dessa (acquired by Square/Block). The team is led by Ragavan Thurairatnam and reports directly to Block co-founder Jack Dorsey.

Project Lead

Ragavan Thurairatnam (@RagavanThurair)

Executive Sponsor

Jack Dorsey (@jack)

Core Researchers

Louay Hazami · Hashiam Kadhim · Qi (Alex) Liu · Rayhane Mama · Aydin Polat · Marc Tyndel

Research Interns

Cara Ji · Jibran Iqbal Shah · Dvir Zagury-Grynbaum

Former Interns

Aryan Sharma · Olivia Yan

Open Call for Audio AI Safety Researchers

Because a model that sounds indistinguishably human and runs locally on consumer hardware introduces novel safety, impersonation, and emotional-attachment risks, Block has not yet released model weights or a public API. Team Dessa is actively collaborating with external safety researchers.

Safety & Research Inquiries: ragavan+dessa@block.xyz

block.xyz/dessa ↗

Frequently Asked Questions About Dessa AI

What is Dessa AI by Block?

Dessa AI is an experimental full-duplex speech-to-speech foundation model developed by a deep learning research lab at Block, Inc., trained from scratch exclusively on raw human conversation audio with zero text.

Does Dessa AI use a text LLM or speech-to-text under the hood?

No. Dessa AI uses no text transcripts, no automatic speech recognition (ASR), no text LLM backbone, and no separate text-to-speech (TTS) vocoder. It processes and generates audio waveforms directly.

How much compute was required to train Dessa AI, and where can it run?

Dessa AI was trained on a single NVIDIA GB200 NVL72 system and runs faster than real-time locally on an Apple MacBook Pro equipped with an M4 Max chip or better.

Where can I listen to all Dessa AI audio samples?

You can stream all 21 unsteered official recordings with synchronized dual-channel waveforms, Moment markers, and full transcripts directly on this site.