Block Unveils Dessa, a Voice AI That Learned to Talk by Listening Alone
Block, the payments company behind Square and Cash App, has unveiled Dessa, a full-duplex speech-to-speech AI model that it says learned to talk entirely by listening to human conversation. According to Block, Dessa was trained from scratch on nothing but recorded human speech, with no text and no pretrained models behind it. The company announced the model on October 23, 2026.
How voice AI usually works, and what Dessa does differently
Most voice assistants today work in stages. They turn your speech into text, run that text through a language model, and then read the answer back out loud as speech. Block argues that this pipeline throws away information that text alone cannot carry, including tone, timing, emotion, and emphasis.
Dessa skips the text step entirely. Block describes it as a true speech-to-speech model: audio in, audio out. It is full-duplex, which means it listens and talks at the same time, much like a person does in a conversation. Because it learned from real conversation, Block says Dessa knows when to cut in and when to backchannel with short acknowledgements such as “yeah!”, “got it!”, and “ooh” to show it is following along.
Block also says Dessa natively hears and speaks emotion, sarcasm, laughter, a frustrated sigh, and a voice cracking with sadness. The company says inference runs faster than real time, so there is no awkward wait for a reply, and that the model runs locally on a MacBook Pro with an M4 Max chip or better.
Learning language from sound alone
The training claim is the core of the announcement. Block writes that, as far as the company knows, Dessa is the world’s first AI you can have a realistic live conversation with that was trained only from human speech.
Block says Dessa learned language from hundreds of times fewer words than large language models need, and that it trained the model on a single NVIDIA GB200 NVL72 system, a far smaller setup than the large clusters typically used for frontier text models. Block has not published a parameter count or a technical paper, so these efficiency claims rest on the company’s own account for now.
To support its claims, Block published recorded conversations that it describes as live recordings of people talking to Dessa, with no one steering the model and subtitles added afterward. In the clips, Dessa laughs mid-sentence when something is funny, gets louder when asked to speak up, and finishes a person’s sentence when the person blanks on a movie title. One recording has Dessa conversing with GPT-Live-1, a third-party AI model; Block notes that GPT-Live-1’s developer is not affiliated with Block and has not reviewed or endorsed Dessa.
Who built it, and who gets access first
Dessa comes from a very small AI research lab inside Block, founded by former members of Dessa, the Toronto deep learning startup. The lab is led by Ragavan Thurairatnam and reports directly to Jack Dorsey, Block’s cofounder, chairman, and Block Head. The team says it has built deep learning models since 2012 and generative speech models for more than eight years.
The lab gives three reasons for building Dessa: curiosity, better understanding between people and AI, and safety. Its safety argument is that today’s voice AI is nearly deaf to the emotion in a speaker’s voice. Block gives the example of a person in serious pain while a voice AI has no idea and just keeps going.
Dessa is not publicly available yet. Block is giving a small group of AI safety researchers early access to find risks the team has not anticipated, and it runs a public waitlist alongside that program. No wider release date has been announced.
What is still unproven
Independent reaction so far has been mixed. The Verge found the demo conversations mostly natural but sometimes “awkward and uncanny.” In one clip, the model starts off thinking the speaker is someone named “Ronaldo.” That reaction matters because the demo clips were chosen by Block, so they show what Dessa can do at its best, not how often it succeeds.
There is also much that Block has not shared: no model size, no benchmark results, no latency figures, no supported languages, no API, no download, and no price. Until outside researchers can test the model, the headline claims remain company claims rather than independently verified results.
Why it matters
If the speech-only approach holds up under independent testing, it could change how voice AI is built. Today’s dominant approach assumes that large amounts of text are needed to teach a model language. Dessa suggests that conversation audio alone might be enough, and that learning from sound could preserve the emotional and conversational cues that text pipelines discard.
That has practical implications for voice assistants, customer support, and accessibility tools, where how something is said often matters as much as what is said. It also raises new safety questions: a voice that sounds convincingly human could be misused for impersonation or phone scams. Block’s decision to give safety researchers first access suggests the company is aware of that risk, even if the details of its safeguards are not yet public.
