Dessa AI is a full-duplex speech-to-speech voice model from a small research lab at Block, the company behind Square and Cash App. Block says Dessa was trained from scratch only on recordings of human conversation, with no text and no pretrained language model.

Who made the Dessa voice model?

A small AI research lab at Block, founded by former members of Dessa, the Toronto deep learning startup Square acquired in 2020. The project is led by Ragavan Thurairatnam and the lab reports to Jack Dorsey.

What does full-duplex mean for Dessa?

Full-duplex means Dessa listens and talks at the same time. It can say 'yeah' while you are still speaking, laugh mid-sentence, or cut in at a natural pause instead of waiting for silence.

Can Dessa run locally?

Block says Dessa runs faster than real time and can run locally on a MacBook Pro with an M4 Max chip or better. This has not been independently tested because the model is not public.

Unofficial guide. This site is not affiliated with, endorsed by, or connected to Block, Inc. "Dessa" is used only to describe Block's model. Official page: block.xyz/dessa.

Dessa AI: Block's voice model trained only on human speech

Q: Can I use Dessa AI right now?

No. As of October 28, 2026 Dessa has no public app, API, download or price. Block is giving early access to a small group of AI safety researchers and says launch links are coming soon.

Dessa AI is a full-duplex speech-to-speech model announced by Block on October 23, 2026. Block says it learned to hold a live conversation by listening to people talk, with no text in training at all. This page collects what is known about the Dessa voice model, what is only claimed, and what you can and cannot do with it today.

TypeSpeech-to-speech, full-duplex

Training dataHuman conversation audio only

Runs onMacBook Pro M4 Max or better (per Block)

Public accessNone yet (Oct 28, 2026)

What is Dessa AI?

Most voice assistants work as a pipeline. Your speech is turned into text, a language model writes a text reply, and a text-to-speech engine reads it out. Dessa skips the text step. Audio goes in, audio comes out. Block describes it as a true speech-to-speech model in which "everything it knows came from human speech."

Block's argument is that the text step throws information away. Tone, timing, a sigh, sarcasm or a voice cracking with sadness do not survive the trip through plain text, so a pipeline model cannot react to them. Dessa AI is meant to hear and produce that emotion directly.

How the Dessa voice model works (what Block has shared)

Full-duplex. Dessa listens and talks at the same time. It backchannels with "yeah!", "got it!" or "ooh" to show it is following, and it can cut in at a natural moment.
Trained from scratch on speech. No text, no pretrained models. Block says Dessa learned language from "hundreds of times fewer words" than a typical large language model.
Small training setup. Block says Dessa was trained on a single NVIDIA GB200 NVL72 system.
Fast and local. Inference runs faster than real time, and Block says the model runs locally on a MacBook Pro M4 Max or better.

Block has not published a technical paper, benchmark results or model weights. Every number above is Block's own claim and has not been checked by anyone outside the company.

Dessa AI demos

Block's announcement page has six live recordings. Block says nobody steered the model and subtitles were added afterward:

Talking about the release. Dessa laughs mid-sentence and speaks louder when asked.
Remembering a movie. The person forgets a movie title and Dessa finishes the sentence.
An anniversary surprise. Dessa gasps "oh my god" and answers with what sounds like teary gratitude.
Expressive conversation. A wide range of intonation and speed.
Life advice. Block describes this clip as empathetic.
Dessa in conversation with GPT-Live-1. Two AI voices talking to each other. Block notes the GPT-Live-1 developer is not affiliated with Block.

Listen to them on the official Dessa page. Early reviews are mixed. MindStudio found some exchanges, such as a laugh that lands inside a sentence, convincing, and other clips stiff or "plasticky".

Can you use Dessa AI?

Not yet. As of October 28, 2026 there is no Dessa app, no Dessa API, no download and no price. Block is giving early access to a small group of AI safety researchers and says "launch links are coming soon." Any website that offers a Dessa download, Dessa API key or "Dessa voice generator" today is not Block's model. See Dessa release date and access for the details we track.

Who built Dessa?

Dessa comes from a very small AI research lab at Block, founded by former members of Dessa, the Toronto deep learning startup that Square acquired in February 2020. The team says it has built deep learning models since 2012 and generative speech models for more than eight years. The lab is led by Ragavan Thurairatnam and reports directly to Jack Dorsey.

Why Block built a speech-only model

Block lists three reasons: curiosity, better understanding between humans and AI, and safety. On safety, Block's point is that today's AI is nearly deaf to emotion in a voice. A voice AI that cannot hear someone is in pain will just keep going. Block also already uses voice AI for ordering in Square, so a model that runs on-device could lower the cost of voice features for businesses, as TechCrunch noted.

Dessa AI FAQ

Is Dessa AI open source?

No weights or code have been released. Block has not said whether Dessa will be open source.

Is Dessa related to Dessa the company?

Yes. The lab was founded by former members of Dessa, the Toronto startup Square bought in 2020. The model shares the name.

How is Dessa different from ElevenLabs or GPT-Live?

ElevenLabs makes text-to-speech, which reads text aloud. GPT-Live is OpenAI's full-duplex voice model. Block's claim is that Dessa is the first realistic conversational model trained only on speech, with no text at any stage.

Does Dessa support languages other than English?

All published demos are in English. Block lists "how efficiently can they learn new languages?" as an open research question, so other languages are not confirmed.

Sources: Block, "Dessa"; TechCrunch, Oct 24, 2026; MindStudio, Oct 25, 2026; Square, Feb 7, 2020. Last updated October 28, 2026.