Block’s AI research lab has released Dessa, a full-duplex speech model trained only on human conversation. Block announced it on October 23, 2026 as a research preview. Dessa is a speech-to-speech model from Block that listens and talks at the same time. Block says it used no text and no pretrained models. That is a different design from the text-first stacks behind most voice agents today.
TL;DR
- Runs on: A MacBook Pro M4 Max or better, locally, faster than real time.
- Performance: Demo-only so far. Live recordings show backchannels, mid-sentence laughs and natural interruptions.
- Best: Trained on 1 NVIDIA GB200 NVL72 rack (72 GPUs). Moshi used 1,016 H100 GPUs.
- Worst: 0 published benchmark scores. No weights, paper, API or price.
What is Block’s Dessa?
Dessa is a true speech-to-speech model: audio in, audio out. According to Block’s project page, everything it knows came from human speech.
It is full-duplex, so it listens while it talks. Block says it learned from real conversation when to cut in. It also learned when to backchannel with short cues like “yeah” or “got it”.
Block’s demo clips show Dessa laughing mid-sentence and getting louder when asked to speak up. One clip shows it talking with OpenAI’s GPT-Live-1. Block says nobody steered any clip.
How does speech-only training work?
Most voice systems chain 3 parts: speech recognition, a text model, then speech synthesis. Block team argues that chain throws away emotion, sarcasm, laughter and sighs. Dessa hears and produces those signals directly.
Block has not published the architecture. Here is what it has disclosed:
- Training data: human conversation audio only. No text and no pretrained models.
- Compute: 1 NVIDIA GB200 NVL72, which links 72 Blackwell GPUs and 36 Grace CPUs.
- Data efficiency: Block says Dessa learned language from hundreds of times fewer words than LLMs need.
- Inference: faster than real time, on a MacBook Pro M4 Max or better.
The lab also explains why this is hard. Speech never repeats exactly. A spoken word takes thousands of times more bytes than the same word in text.
Kyutai’s Moshi, the best-known open full-duplex model, takes another route. Moshi starts from a text language model backbone and predicts time-aligned text tokens before audio tokens. Its model card lists 7 million hours of pretraining audio and 1,016 H100 GPUs.
‘;
h+=”});
$(‘#md-flow’).innerHTML=h;
$(‘#md-chips’).innerHTML=f.chips.map(function(c){return ‘‘+c+’‘}).join(”);
var cs=$$(‘#md-chips .chip’);setTimeout(function(){cs.forEach(function(c,i){if(i>=f.keep)c.classList.add(‘lost’)})},450);
$(‘#md-note’).innerHTML=f.note;setTimeout(post,60);}
$$(‘.seg button’).forEach(function(b){b.onclick=function(){$$(‘.seg button’).forEach(function(x){x.classList.toggle(‘on’,x===b)});drawFlow(b.getAttribute(‘data-m’))}});
drawFlow(‘c’);
// duplex
var N=48;[‘#md-w1′,’#md-w2’].forEach(function(id){var w=$(id);for(var i=0;i
var on2=s.d.some(function(r){return i>=r[0]&&i<=r[1]});w2[i].style.height=(on2?6+Math.abs(Math.sin(i*1.3+tick*0.9))*34*s.dl:4)+’px’;}
$(‘#md-w1’).classList.add(‘live’);$(‘#md-w2’).classList.add(‘live’);}
function scene(k){cur=k;var s=scenes[k];$(‘#md-cap’).textContent=s.cap;$(‘#md-ps’).textContent=s.ps;$(‘#md-ds’).textContent=s.ds;$(‘#md-snote’).textContent=s.note;
$$(‘#md-steps button’).forEach(function(b){b.classList.toggle(‘on’,+b.getAttribute(‘data-s’)===k)});render();setTimeout(post,60);}
$$(‘#md-steps button’).forEach(function(b){b.onclick=function(){scene(+b.getAttribute(‘data-s’))}});
scene(0);
var reduce=window.matchMedia&&matchMedia(‘(prefers-reduced-motion: reduce)’).matches;
if(!reduce)setInterval(render,180);
// bars
function bars(){$$(‘.fill’).forEach(function(f){f.style.width=”0″;setTimeout(function(){f.style.width=f.getAttribute(‘data-w’)+’%’},80)})}
// disclosed grid
var cells=[
[‘Input / output’,’Audio in, audio out’,1,’True speech-to-speech; full-duplex, listens while talking.’],
[‘Training data’,’Human conversation only’,1,’No text and no pretrained models, per Block.’],
[‘Training compute’,’1 GB200 NVL72′,1,’One rack: 72 Blackwell GPUs and 36 Grace CPUs (NVIDIA spec).’],
[‘Local inference’,’MacBook Pro M4 Max+’,1,’Faster than real time, per Block. X post from the lead says consumer GPUs.’],
[‘Parameters’,’Not disclosed’,0,’No size, context length or architecture published.’],
[‘Hours of audio’,’Not disclosed’,0,’Block says only that it used hundreds of times fewer words than LLMs.’],
[‘Benchmarks’,’None published’,0,’All evidence so far is live recordings chosen by Block.’],
[‘Weights / API / price’,’Not available’,0,’Research preview with a waitlist; safety researchers get early access.’]];
$(‘#md-grid’).innerHTML=cells.map(function(c){return ‘
‘+c[0]+(c[2]?”:’missing‘)+’
‘+c[1]+’
‘+c[3]+’
‘}).join(”);
$$(‘.cell’).forEach(function(c){c.onclick=function(){c.classList.toggle(‘open’);setTimeout(post,380)}});
window.addEventListener(‘load’,post);window.addEventListener(‘resize’,post);setTimeout(post,300);
})();
