This site uses cookies for analytics and to improve your experience. By clicking Accept, you consent to our use of cookies. Learn more in our privacy policy.

Skip links
AIWIZ Digital Marketing

What is Dessa AI? Block’s speech-only voice model explained

Author AIWIZ Digital
Published on:
Published in:
Dessa in pale grey on near-black, between two timelines of speech segments that overlap in lime
Block says Dessa learned to hold a conversation from recordings of human speech alone.

The answer in brief

  • Dessa is a voice AI model from Block, the company behind Square and Cash App. Block announced it on 23 October 2026.
  • Block describes it as a speech-to-speech model: audio goes in and audio comes out, with no text step in between.
  • Block says it was trained from scratch on recordings of human conversation alone: no text and no pretrained models.
  • It is full-duplex, so it listens while it talks. In Block’s recordings it laughs mid-sentence, cuts in and drops in a “yeah” while the other person speaks.
  • Block says it runs locally on a MacBook Pro with an M4 Max chip or better, and was trained on a single NVIDIA GB200 NVL72 system.
  • As of 28 October 2026 you cannot use it. Dessa is a research preview: a small group of AI safety researchers gets early access first, and there is no API, price, model card or download.
  • Sites already selling “Dessa AI” plans are not Block’s model. Their live calls run on OpenAI’s GPT-Live.
  • Our view: the demos are impressive, but every clip is Block’s own. Until outside researchers report back, treat Dessa as a signal of where voice AI is heading. It is too early to plan around.

What is Dessa AI?

Dessa AI is a research preview of a conversational voice model built by a small AI research lab inside Block. Block’s page calls it a conversation model trained only on speech, and its pitch is simple: you learned to talk by listening, and so did Dessa.

Jack Dorsey, Block’s co-founder and chairman, introduced it on X on 23 October 2026 as “the most authentic full-duplex speech model ever built”. The lab’s lead, Ragavan Thurairatnam, went further in his own post, calling it a GPT-3 moment for speech and saying it can run on consumer GPUs.

The name causes confusion. “Dessa” is the 2026 voice model, and it was also the Toronto deep learning startup Square bought in 2020, whose former staff founded the lab. Trengo, a customer-service software company, sells an unrelated e-commerce AI agent called Dessa, and two paid voice studios use the name with no link to Block (more on those under “Can you use Dessa yet?”). This post is about Block’s model; the startup’s story is further down.

How is Dessa different from other voice AI?

Dessa, Block says, never turns speech into text. Older voice assistants run a chain: speech recognition writes down what you said, a text model works out a reply, and a speech engine reads that reply aloud. Block’s complaint is that the written step throws away tone, sarcasm, laughter, a sigh or a voice cracking with sadness. Its page puts it bluntly: “Interacting with current voice AI can just plain suck.”

Three claims set Dessa apart from that chain:

  • Speech only, from scratch. Block says every fact Dessa knows came from raw sound. No text data and no pretrained model were used, and it claims Dessa learned language from hundreds of times fewer words than a large language model needs.
  • Full-duplex. It listens and speaks at the same time. Block says it learned from real conversations when to interrupt and when to backchannel with short sounds such as “got it” or “ooh”.
  • Small enough to run locally. Inference runs faster than real time, and Block says a MacBook Pro M4 Max or better can run it. Training took a single NVIDIA GB200 NVL72 system: one rack, but one holding 72 Blackwell GPUs and 36 Grace CPUs.

The training is what sets Dessa apart, because full-duplex speech is no longer unusual. OpenAI’s GPT-Live models, first released on 8 July 2026, also listen and talk at once. SpaceXSI (formerly SpaceXAI) sells Grok Voice 3.0, released on 14 October 2026, as a speech-to-speech model, though it has not published how that model is built or trained.

Two claims are easy to blur here. A model can work on audio directly while it talks, as GPT-Live-1 and Kyutai’s Moshi do, and still lean on text. GPT-Live-1 hands harder work to a text model, and Moshi is built on a 7-billion-parameter text model called Helium. Block’s claim covers both halves. Dessa runs on speech and, Block says, learned only from speech.

Three voice AI routes: older assistants turn speech into text, GPT-Live adds a backend model, Dessa uses one speech model
Figure 1. How three kinds of voice model handle your speech. Dessa’s route is as Block describes it; GPT-Live’s backend model is from OpenAI’s model page. Checked 27 October 2026.

Block’s “world’s first” also needs reading closely. Meta’s dGSLM, published in 2022, was trained on 2,000 hours of two-channel phone conversations with no text or labels, and produced laughter, overlapping speech and natural turn-taking. It generated both sides of a dialogue by itself, though, and never talked live with a person. Its authors said it had not seen enough data to capture the deeper syntax and meaning of conversation, and Kyutai’s Moshi paper later called it a proof of concept that could not run online. Block’s claim is narrower, and still new if it holds up: a live conversation partner, holding up its end of a real exchange, that learned only from speech.

Block gives three reasons for the project: curiosity, better understanding between people and AI, and safety. The safety argument is that a voice assistant which cannot hear pain or distress will carry on regardless, and that an AI which hears emotion natively is a step towards fixing that.

What do the Dessa demos show?

The demos show a model with an unusually human sense of conversational timing. Block’s page leads with six described recordings, followed by a “More samples” section, and says each is a live conversation that nobody steered. Dessa chose what to say, when, and in what voice and personality. Subtitles were added afterwards.

The moments worth listening for:

  • Laughing on time. Dessa laughs the moment something is funny, mid-sentence, where other models wait for you to finish. Asked to speak up, it gets louder.
  • Finishing a thought. A speaker forgets a film title and Dessa supplies it, which suggests it has picked up a fair stock of facts from audio alone.
  • Emotional range. Told about an anniversary surprise, it gasps under its breath and replies with what sounds like teary gratitude. Another clip gives life advice in an empathetic tone.
  • Talking to another AI. One recording has Dessa in conversation with GPT-Live-1. Block labels both voices as AI-generated and notes that OpenAI is not affiliated with Block and has not endorsed Dessa.

Block and its researchers have posted more clips on X, including a guessing game with intern Cara Ji. Gizmodo spotted a slip in one of the guessing-game clips: Dessa carried on guessing after being told it had won. It is a useful reminder. The voice sounds human; whether the answers are reliable is a separate question the demos cannot settle.

Who built Dessa?

Former members of the Toronto startup that made a Joe Rogan voice deepfake in 2019, though this is a different model. Dessa comes from a very small research lab at Block, led by Ragavan Thurairatnam and reporting directly to Jack Dorsey. Block says the team has built generative speech models for more than eight years.

That startup began in Toronto in 2016 as DeepLearni.ng, founded by Stephen Piron, Ragavan Thurairatnam, Eric Kin-Ho Lee and Vincent Wong. It raised a $9 million Series A in September 2017 and renamed itself Dessa in June 2018, when it also signed AlexNet co-author Alex Krizhevsky as a technical adviser.

In 2019 Dessa engineers released RealTalk, a synthesised Joe Rogan voice close enough to the real thing to make headlines. Two of the three people credited with that demo, Hashiam Kadhim and Rayhane Mama, appear on the 2026 Dessa team list.

Square, which renamed itself Block in 2021, announced it had bought Dessa on 7 February 2020 for an undisclosed sum. Square’s announcement noted that many people knew Dessa for its work on deepfake detection. Block’s new page does not mention RealTalk, and nothing Block has published presents the new model as RealTalk’s successor.

The full team as Block lists it: researchers Louay Hazami, Hashiam Kadhim, Qi (Alex) Liu, Rayhane Mama, Aydin Polat and Marc Tyndel; interns Cara Ji, Jibran Iqbal Shah and Dvir Zagury-Grynbaum; and former interns Aryan Sharma and Olivia Yan.

How does Dessa compare with OpenAI’s GPT-Live?

OpenAI’s GPT-Live is a product you can use today; Dessa is a research result you can only listen to. Block’s demo talks to GPT-Live-1, but ChatGPT began moving to GPT-Live-1.1 on 15 October 2026, while OpenAI’s API still offers gpt-live-1. Both are full-duplex, so the comparison that matters is where each gets its knowledge and how you can get at it.

Dessa against OpenAI’s full-duplex voice models

If the table extends beyond the screen, scroll sideways to view all columns.

Area Dessa (Block) GPT-Live (OpenAI)
Announced Dessa (Block)23 October 2026 GPT-Live (OpenAI)8 July 2026 (GPT-Live-1); 15 October 2026 (GPT-Live-1.1)
Status Dessa (Block)Research preview, early access for safety researchers GPT-Live (OpenAI)GPT-Live-1.1 in ChatGPT Voice; gpt-live-1 in the API
Listens while speaking Dessa (Block)Yes GPT-Live (OpenAI)Yes
Where its answers come from Dessa (Block)Everything it knows was learned from speech GPT-Live (OpenAI)Hands reasoning and tool use to a backend model
Where it runs Dessa (Block)Locally, on a MacBook Pro M4 Max or better GPT-Live (OpenAI)OpenAI’s servers
Price Dessa (Block)None published GPT-Live (OpenAI)gpt-live-1: $0.05 a minute in the API, backend model billed separately (US list price)

Sources: Block’s Dessa page, OpenAI’s GPT-Live 1 model page and ChatGPT release notes, checked 27 October 2026.

The design split is the interesting part. OpenAI keeps the voice layer light and lets a stronger text model do the thinking, which is how GPT-Live can take on harder questions and call tools without stalling the conversation. Dessa bets the other way, on one model that hears and speaks everything itself. Block argues that makes it better at tone and timing. It may also leave it weaker on facts, and Block has published nothing yet to show either way.

Can you use Dessa yet?

No. There is no app, API, price or downloadable model, and Block has given no release date. You can listen to the recordings on Block’s Dessa page without signing up.

For now there is one route in, and one place to watch:

  • AI safety researchers can ask for early access through a form on Block’s page, which asks for an institutional email from a university, company or lab.
  • Everyone else should watch Block’s page. Some outlets reported a public waitlist at launch, but Block has not opened one: its page says only that launch links are coming soon.

Be careful with sites that sell “Dessa AI”. At least two paid voice studios use the name, and neither runs Block’s model:

  • dessaai.net calls itself an independent service, not affiliated with Block. Its live calls run on OpenAI’s GPT-Live-1 and its text to speech on OpenRouter.
  • dessa-ai.com advertises plans from $13.90 a month, billed yearly, in US dollars. Its own FAQ says that, because Block’s model has no public API yet, its live calls run on GPT-Live-1 and its text to speech on HeyGen Voice.

Both sites feature recordings of Block’s model, which makes the mix-up easy. A live call on either is a call with OpenAI’s model. The official home is block.xyz/dessa, and new recordings are due on Thurairatnam’s X account.

What has Block not said about Dessa?

A lot of what you would need to judge Dessa is missing. Block’s page and launch posts give no:

  • model size, benchmark scores or latency figures (CoinDesk also noted the first two were missing);
  • model card, red-team results or safeguards against voice cloning;
  • details of the training audio: where the recordings came from, how many hours there were, or whether the speakers agreed to their use;
  • list of the languages it speaks;
  • release date, price or licence.

TechCrunch raised one more open question. Rival speech engines such as ElevenLabs v4 steer emotion with written tags, and it is not clear whether anything similar sits behind Dessa’s expressiveness. Block says Dessa hears and speaks emotion natively, but no independent tests have been published yet.

None of this makes the claims false. It does mean the samples are, as CoinDesk put it, “company demonstrations rather than independent tests”. The safety researchers now getting access are the first people who can tell us how Dessa behaves when Block is not choosing the clips.

Should businesses worry about Dessa and voice scams?

Yes, though Dessa itself is not the threat today. The reaction to the launch split quickly. One listener on X said he had never heard anything artificial sound so human, and Binance founder Changpeng Zhao saw a use in teaching children. Others went straight to fraud: journalist Shiv Aroor predicted scam calls would reach pandemic levels, and filmmaker Justine Bateman asked why an AI voice needs to sound human at all.

The worry is fair. A voice that laughs, hesitates and sounds upset at the right moment removes the cues people use to spot a robocall. Dessa running on a laptop sharpens the point: if a model like it were ever released as downloadable weights, nobody could switch off misuse from a central server, the way an API provider can.

Block has a stake in getting this right. It runs Cash App and Square, so convincing fake voices are a fraud risk to its own customers, and putting safety researchers ahead of developers is a sensible first step. Whatever happens to Dessa, the practical defences for any business stay the same:

  • confirm payment or bank-detail changes through a second channel you already trust, never on the call that asked for them;
  • agree verbal passphrases for urgent requests between senior staff and finance;
  • train staff that a natural, emotional voice proves nothing about who is speaking.

What does Dessa mean for UK businesses?

For now, nothing you can buy, and nothing to change in this quarter’s plans. If you are building a voice agent for bookings, support or sales, test the models that are shipping: OpenAI’s GPT-Live, Grok Voice 3.0, Google’s Gemini 3.8 Live and Mistral’s Voxtral 3. Dessa cannot be piloted, priced or put through a data protection review.

It still matters as a direction of travel. Block has not said Dessa will reach its products, but Square already offers AI voice ordering to US sellers, and, as TechCrunch pointed out, a model that runs on the device would make features like that much cheaper to offer. Running locally would not settle data protection questions on its own, though: the application around a model can still log audio or send it elsewhere. If speech-only models deliver what the demos suggest, customers will come to expect a voice agent that notices when they are annoyed.

Whichever voice model you trial, Dessa if and when it arrives or a shipping model today, judge it on the same points:

  • callers are told at the start that they are speaking to AI;
  • there is a tested route to a person, and it works when the agent is confused or unavailable;
  • it responds sensibly when a caller sounds distressed or asks for something it should not do;
  • you know what audio and transcripts are kept, where, for how long and who can see them;
  • it copes with UK regional accents, local place names and surnames, background noise and hesitant speakers;
  • people who cannot or would rather not use voice have a text route;
  • success is measured in tasks resolved and corrections handled, because a human-sounding voice proves nothing by itself.

Our view: Dessa is a striking research result that the public cannot use yet. The claim that a model can learn to hold a conversation from sound alone, on one rack of hardware, is a genuine surprise if it holds up. Wait for the safety researchers’ findings, and for a model card, before deciding how much of it does.

If you want help working out where voice AI fits in your customer journeys, talk to AIWIZ.

Sources

Explore
Drag