Dessa AI: Block's voice model that never saw a word of text

On October 23, Block, Jack Dorsey’s company, unveiled Dessa AI, a speech-to-speech voice conversation model trained from scratch purely on recordings of human speech. No text, and no pre-trained model to start from. The announcement made the front page of Hacker News, where it collected almost 190 points and a few dozen comments.

What’s new about Dessa AI?

Most voice assistants work like this: speech gets turned into text, a language model replies in text, and at the end that text is turned back into speech. Along the way you lose emotion, irony, laughter and pauses.

Dessa AI skips that step. According to the announcement:

1. It has never seen a single word of text – it learned only from recorded human conversations. The Dessa team compares this to how babies learn to talk before they ever see a written word (quoting “Babies learn to talk without ever seeing a word written down”).
2. It works in full duplex, meaning it listens and speaks at the same time. It can drop in a “yeah” or “got it”, interrupt at a natural moment, and react to a sigh or a breaking voice.
3. It picked up language from hundreds of times fewer words than large language models. Block didn’t give an exact number.
4. Training ran on a single NVIDIA GB200 NVL72 system.
5. It runs faster than real-time text models and can be run locally on a MacBook Pro with an M4 Max or better.

Granted, according to Block a single spoken word takes thousands of times more bytes than the same word written as text - which makes it all the more surprising that one rack was enough for training.

Who is behind Dessa?

A very small team: six researchers and a few interns. It’s led by Ragavan Thurairatnam, co-founder of the Toronto startup Dessa, which Block acquired in 2020. According to AI Polix, the lab reports directly to Dorsey, and the team has been building generative speech models for more than eight years.

What do conversations with Dessa sound like?

Block published captioned recordings: Dessa helps someone recall the title of a film they can’t remember, gets emotional at news of an anniversary surprise, gives life advice, sings in Spanish and plays guessing games. There’s also a series of conversations with GPT-Live-1, OpenAI’s voice model.

On Hacker News, some people are impressed by how fluid the conversation is (one commenter writes that it’s “clearly far far superior in terms of fluidity of conversation”), but others point out that it sounds great while there isn’t much “smarts” behind it: “The intelligence sounds lacking”. The Verge also describes the conversations as mostly natural but awkward at times. In one recording the model greets the other person as Ronaldo ;)

What to keep in mind

First, all of this comes from Block’s own claims. There’s no research paper, no benchmarks, no model size and no license, and the demo was carefully curated. Let’s treat it with some skepticism until independent opinions come in.

Second, safety. A voice that sounds human is also a ready-made tool for scammers, and that’s the worry that comes up most often in the comments. Block argues the opposite: according to the team, today’s AI is “nearly deaf to the emotion in our voices”, so it can miss a person in crisis, while a model that hears emotion could respond better.

How to try Dessa AI?

For now, you can’t. On Dessa’s page you can join the waitlist, and AI safety researchers can request early access using their institutional email address. There’s no date for a wider release, and it’s unknown whether the weights will be open.

If Dessa really does land on laptops, it’ll be one more model you can run on your own machine, which I covered in my guide on how to run AI locally. For comparison, a ready-to-use voice API is, for example, Mistral’s Voxtral 3, which however works in the classic transcription-based setup.

FAQ

What is Dessa AI?

It’s an experimental voice conversation model from Block (Jack Dorsey’s company). It takes in and generates audio directly, without converting speech to text, and it can listen and speak at the same time.

Is Dessa AI available for free?

It isn’t publicly available yet. Block is giving access only to selected AI safety researchers and runs a waitlist. No pricing or launch date has been announced.

Does Dessa speak languages other than English?

Block hasn’t said. The published recordings are mostly in English, plus one with singing in Spanish.