For years, voice AI has followed the same pipeline: transcribe the speech into text, hand that text to a language model, and translate the reply back into audio. Everything the pipeline cannot carry — the tone, the pause, the sarcasm, the sigh — gets flattened on the way in.
On October 23, Block's in-house AI research lab published a model that runs the route backwards. It is called Dessa, and it is a full-duplex speech-to-speech model: audio in, audio out, with no text stage anywhere in between. The project page lives at block.xyz/dessa.
What exactly is different
Block's description is unusually blunt: the model was trained from scratch on recorded human conversation only. No text, no pretrained models. It is genuinely full-duplex, so it listens and talks at the same time the way a person does. It knows when to cut in, and when to backchannel with a short "yeah!" or "got it!" to signal that it is still following.
In the published demos, it laughs mid-sentence when something is funny. When a speaker blanks on a movie title, it finishes the sentence for them. Ask it to speak up and it gets louder. Block says every clip is a live recording that nobody was steering, with subtitles added afterward.
Training on sound alone turns out to be far harder than it sounds. Block points out that one spoken sentence carries thousands of times more bytes than the same sentence as text, and that speech never repeats: say a line twice and the waveform still comes out different. Babies learn to talk without ever seeing a word written down, and Block wanted to know whether a model could do the same.
The answer appears to be yes, but the training bill was one single NVIDIA GB200 NVL72. One system, not a cluster.
The interesting part is not "first"
The project page contains a sentence that Block itself qualifies with "as far as we know": this is the first AI you can have a realistic live conversation with that was trained only from human speech. Only Block is making that claim, and no third party has reproduced it, so it is better treated as an open question than as a settled result.
The more consequential number is a different one. Block says Dessa learned language from hundreds of times fewer words than text-based models. Crypto Briefing's report adds the caveat that the company did not provide a detailed comparison.
If the direction holds, "is speech a cheaper carrier than text" gets recalculated from scratch. The industry has spent years compressing tokens and hunting for cheaper representations, always with the same premise: text is expensive. Dessa attacks the premise from the other side. Maybe the signal humans actually use to talk was never written language in the first place.
No weights, no benchmarks, only demos
This is where a cooling-off moment is warranted. So far Block has not disclosed a parameter count, dataset size, latency figures, or any formal benchmark results. What is public is a set of demo recordings that Block selected itself, and the model's performance has not been independently evaluated.
Independent coverage noticed the same gaps. Crypto Briefing noted that some demonstrations show the model misidentifying a speaker or unexpectedly changing topics. AI Chat Daily likewise describes the release as a research preview rather than a product, with early access limited to a small group of AI safety researchers who must apply with an institutional email.
That ordering is telling. Safety researchers first, users later. A model that learned to read emotion from scratch raises the question of what risks nobody anticipated before it reaches a product surface.
Who is behind it
The team calls itself Team Dessa and is led by Ragavan Thurairatnam. Most of its members came from Dessa, the Toronto speech-AI startup, and the model inherited the name. Block says the lab has been building deep learning models since 2012 and generative speech models for more than eight years, and reports directly to Jack Dorsey.
Block gives three reasons for the project: curiosity, better understanding between humans and AI, and safety. The last two are really the same thing. If an AI can hear pain in a voice, it should not keep mechanically talking while someone is in distress. The team's own line is that AI keeps gaining power and independence while remaining nearly deaf to the emotion in our voices.
Early days, but the question is real
There is no public weight release and no fuller launch timetable yet. The comparatively solid facts are that inference runs faster than real time and works locally on a MacBook Pro with an M4 Max chip or better, which three independent sources repeat.
So it is too early to draw conclusions. What Dessa does is overturn an assumption the industry stopped questioning years ago: voice AI goes through text because there was no other way to teach a model to hold a conversation. That sentence now needs another look.