What is Dessa, Block's full-duplex voice model?

What Dessa is

Dessa is a conversational voice model. Audio goes in and audio comes out, with no text step in between. Block describes it as a research preview, and Jack Dorsey shared it on X on 2026-10-23.

Block calls Dessa "the most authentic full-duplex speech model ever built." That is Block's own description. Block has not published benchmarks, so nobody outside the company can check it yet.

This page sums up what is known and what is not, with sources. We are an independent company and are not affiliated with Block.

Sources: Block, TechCrunch

What full duplex means here

Most voice assistants take turns. They wait until you stop talking, then reply. A full-duplex model listens and speaks at the same time, the way two people do on a phone call.

Block's main argument is about text. In a typical voice assistant, your speech becomes text, a language model writes a reply, and another system reads it aloud. Block says the trip through text throws away tone and timing, which is why so much voice AI feels slow and flat.

Here is what Block says Dessa can do:

• It makes short listener sounds while you talk, to show it is following.
• It decides when to cut in, instead of waiting for silence.
• It laughs in the middle of a sentence when something is funny.
• It speaks louder when you ask it to, which Block says nobody taught it.
• It hears and expresses emotion, sarcasm, sighs and laughter.

Sources: Block

How Block says it was built

Block is the only source for the details in this section. None of them has been checked by outside researchers.

Block explains why training on speech alone is hard. A spoken word takes far more data than the same word in text, and the same sentence never sounds exactly the same twice. A model with no text to lean on has to learn words, accents, pauses and emotion from raw sound.

Here is what Block says about the model itself:

• Data: trained from scratch on recordings of human conversation, with no text and no pretrained models.
• Words: Dessa learned language from hundreds of times fewer words than large language models use. Block has not said how many hours of audio that was.
• Training compute: one NVIDIA GB200 NVL72. NVIDIA describes that system as a single liquid-cooled rack with 72 Blackwell GPUs and 36 Grace CPUs.
• Inference: Dessa runs locally on a MacBook Pro with an M4 Max chip or better, faster than real time. Block gives no latency figure.
• Size: Block has not published the number of parameters.
• Open questions: Block asks how efficiently models like this learn language and facts compared with text models, and how well they learn new languages.

Sources: Block, NVIDIA

What the demos show

Block's Dessa page has 21 recorded conversations. Block says each one is a live recording of a person talking to Dessa, with nobody steering it, and that subtitles were added afterwards. Topics include a forgotten movie title, an anniversary surprise, life advice, video games and a refund.

Two clips pair Dessa with GPT-Live-1, OpenAI's full-duplex voice model. Block notes that both voices in those clips are AI-generated, and that the developer of GPT-Live-1 is not affiliated with Block and has not endorsed Dessa.

TechCrunch described a model that laughs, pauses and even cries, and pointed to a guessing game with a member of the Block team. GIGAZINE described a clip in which Dessa helps a speaker remember the name of a Zelda game. Gizmodo's writer found ChatGPT's voice robotic next to Dessa in a side-by-side clip.

The company chose these clips. They show what Dessa can do on a good take, not how often it does it, how it copes with a bad line or how accurate it is on facts.

Block’s launch film and five of the recordings are below, as Block published them. The buttons under each recording jump to the moments Block marked in the video.

Sources: Block, OpenAI, TechCrunch, GIGAZINE, Gizmodo

Who built it

Dessa comes from what Block calls a very small AI research lab inside the company. The lab reports directly to Jack Dorsey, Block's co-founder and chairman, and is led by Ragavan Thurairatnam. Block's Dessa page lists six researchers and several current and former interns.

The lab was founded by former members of Dessa, a Toronto deep learning startup that was once called DeepLearni.ng. Square, the company now called Block, announced on 2020-02-07 that it had acquired the startup. Square said at the time that the startup was known for its work on detecting deepfakes, and BetaKit named Thurairatnam as its head of machine learning.

The team worked on synthetic voices long before this model. In 2019, three of its engineers released RealTalk clips that imitated the voice of podcaster Joe Rogan, and the company kept the model private because of the risk of misuse. Two of those engineers, Hashiam Kadhim and Rayhane Mama, are on Block's list of Dessa researchers.

Block says the team has built deep learning models since 2012 and generative speech models for more than eight years.

Sources: Block, Square, Gizmodo, BetaKit, The Register, The Batch, DeepLearning.AI

What is not known yet

Some practical questions are open. Voice assistants built on text models usually rely on that model for facts and for calling tools. Block says everything Dessa knows came from the conversations it was trained on, and TechCrunch notes that speech models are usually paired with a large language model for that job. So it is unclear how Dessa would answer factual questions or take actions inside an app.

Safety is the other open question. Commentators quoted by CoinDesk worried that voices this human could make scam calls harder to spot. Block's answer so far is to give AI safety researchers early access to look for risks.

As of 2026-10-31, Block has not published any of the following:

• an API, SDK or price
• model weights, a download or a license
• a paper, model card or benchmark results
• the parameter count or the amount of training audio
• a list of supported languages
• latency figures
• a release date

Sources: Block, TechCrunch, CoinDesk

How to get access

Right now, only AI safety researchers can ask for access. Block says it is giving a small group of them early access to find risks the team has not anticipated. The request form on Block's Dessa page asks for a name, an institutional email address and an optional Google Scholar or blog link.

Everyone else has to wait. When we checked on 2026-10-31, the page's Download button showed only a coming-soon notice, the public waitlist had not opened, and the page marked its launch links as coming soon. Block has not said what will be downloadable, who gets access after the researchers, or when.

Block says more recordings will be shared on X soon.

Sources: Block

What it means for developers

For now there is nothing to call, download or license, so you cannot put Dessa in a product today.

If Block's claims hold up, Dessa still matters. It would show that a voice model can learn conversation from audio alone, and that such a model can run on a high-end laptop. Both points affect the cost, privacy and feel of voice apps.

Teams that need full-duplex voice now have other options. OpenAI sells GPT-Live-1 through its API, and Kyutai's Moshi has open weights you can host yourself. Whatever model you choose, the production work is the same: browser audio over WebRTC, phone calls over SIP, low latency, scaling and monitoring.

That production layer is what we are building. We plan to host Dessa only if Block releases it under a license that allows it, and we will host other full-duplex models either way.

Sources: OpenAI, Kyutai on GitHub