Block Dessa Voice Model Learned to Talk From Speech Alone

Block, the company behind Square and Cash App, released a voice model on Friday, October 23, 2026, that learned to talk by listening to people. The Block Dessa voice model takes sound in and sends sound out, with no written text anywhere in the middle. Most voice assistants today turn speech into text, hand that text to a language model, and then read the answer aloud. Dessa skips those steps, according to Block's research lab.

The lab says it trained the Block Dessa voice model from scratch on human conversation and nothing else. No written text went into the training, and no existing language model sits underneath it. Block says the model learned language from hundreds of times fewer words than large language models see, because real conversation carries things text drops: timing, tone, laughter, and the small sounds people make while someone else is talking.

What the demos actually show

Block published sample conversations instead of a paper or a benchmark. In one demo, the Block Dessa voice model laughs in the middle of a sentence. In another, the voice drops to a quiet aside. When the person on the other end asks it to speak up, it gets louder. One demo has Dessa saying it feels tired after work, and the delivery sounds tired too. The lab's point is that these qualities come from the training data itself, since the model learned from recordings of people talking.

Full-duplex is the technical term for what makes that possible. The Block Dessa voice model listens while it talks, so a person can interrupt it, and it can cut in at natural moments. It also backchannels, offering the small murmured acknowledgements that keep a conversation moving. Ragavan Thurairatnam, who leads the research effort, called the release a GPT-3 moment for speech on X. That label is a claim about importance, and the lab published no measurement behind it.

A laptop instead of a data center

The other headline is where the model runs. Block says inference runs faster than real time on a MacBook Pro with an M4 Max chip or better. Training took a single rack of Nvidia graphics hardware, a small footprint next to the clusters behind frontier text models. Jack Dorsey, who founded Block, posted the announcement on October 23 and wrote that talking with Dessa feels like talking with a person, full of emotion, nuance, pauses, and a natural cadence. When Christopher Stanley asked that evening what language model powers the voice, Dorsey replied that there is none: speech to speech, with no language model involved.

That design also shapes the safety argument Block makes. The lab argues that voice assistants which cannot hear emotion miss information that matters, and that gap can be risky as well as annoying. Early access goes to a small group of AI safety researchers first, through a form on the release page, while everyone else joins a public waitlist. According to Gnosis, Block has not posted the model weights, and the release page gives no benchmark results, release date, or price.

Why skeptics are holding back applause

Other voice systems already do parts of this. ElevenLabs and Fish Audio build emotion into their voices with style tags, a method Block criticizes as bolted on. Kyutai's Moshi handled live two-way speech in 2024, though it started from a text model. OpenAI released a full-duplex voice model earlier this year. Dessa's distinction, according to Block, is that the Block Dessa voice model learned everything from speech alone, which would make it the first realistic live conversation model built that way, as far as the lab knows.

The gap between the claim and the proof is what to watch. Without benchmarks or public weights, outside researchers cannot check how well the Block Dessa voice model handles hard questions, how it copes with languages beyond the demos, or where it fails. The samples show a voice that sounds human. They do not show a system that reasons like one, and Block has not claimed otherwise. The waitlist, not a download button, is what the public gets for now.

Still, the direction matters for anyone building voice products. If a model can learn conversation from audio alone and run on a laptop, the cost and privacy math changes: less data center time, and conversations that never leave the device. Shops and clinics that use voice today mostly rent cloud models. A local model that interrupts gracefully and hears a laugh for what it is would be a different kind of tool. Whether the Block Dessa voice model becomes that tool depends on what Block releases next, and the lab has said little about timing. SiliconNoon reported that the announcement came with sample conversations and a waitlist, and with no paper, benchmark, or API attached, according to SiliconNoon.

According to our earlier reporting, the arrival of AI agents in bank accounts and OpenAI's offline training decision point to the same shift toward AI that runs closer to the user.