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

Block wants its new voice AI to sound less like a machine. Jack Dorsey's company unveiled a voice model named Dessa from its research lab, TechCrunch reported on Oct. 24, 2026.

Block said Dessa was trained entirely on human speech, using a single NVIDIA GPU rack. The model is not open to the public yet.

Key facts

• Block's research lab released Dessa, TechCrunch reported on Oct. 24, 2026.
• Block said the model was trained fully on human speech, on one NVIDIA GPU rack.
• Dessa is full-duplex, so it can listen and speak at the same time.
• Block said it runs locally on consumer computers such as a MacBook M4 Max or better.
• There is no public access to Dessa right now, TechCrunch said.

How Dessa works

Most voice assistants take turns: they wait for you to stop, then answer. TechCrunch said Dessa works in full-duplex mode, so speaking and listening happen together.

Its knowledge also comes from a different place. TechCrunch said speech models usually rely on a paired large language model for knowledge, but Block says Dessa draws its knowledge from its conversation data.

Block argues that text is the problem with many current systems. The company says voice models can sound stiff because text in the pipeline drives how speech is generated, TechCrunch reported. Block summed up its pitch in a blog post: "The most powerful AIs in the world can't hear how you feel."

Emotion is a key part of the claim. Rival models such as ElevenLabs v4 and Fish Audio Drama 3 use voice tags to set emotion, while Block says emotion is built into Dessa. TechCrunch said it is unclear whether any tags steer Dessa.

Block shared samples of the model at work. Its researchers posted example conversations in which Dessa switched voices, pacing and emotions, at times sounding tired or emotional.

What the demos show

Block's samples include some surprises. The Verge said Dessa lead Ragavan Thurairatnam posted a sample chat in which the model at first takes the speaker for someone called "Ronaldo".

Block also staged a head-to-head. One demo pits Dessa against OpenAI's GPT-Live-1 voice model, and The Verge said Dessa sounds more natural in the exchange.

The team gave its reasons for the project. The Verge reported three: curiosity, better understanding between people and AI, and safety. On safety, the team said today's voice AI is nearly deaf to emotion and can miss when a user is in serious pain.

Who is behind it

Block describes the Dessa team as a small AI research lab started by former members of Dessa, the Toronto deep learning startup, and says it reports directly to Jack Dorsey.

The team is making big claims. Dessa lead Ragavan Thurairatnam called the release a GPT-3 moment for speech in a post on X, TechCrunch reported.

What it means for you

For now, you cannot try it. TechCrunch said Dessa has no public access at the moment.

The hardware claim stands out. Block says the model can run on a consumer-grade computer like a MacBook M4 Max or better. Running a voice model on a laptop, rather than in a remote data center, is a different approach from many cloud voice tools.

There is a business angle, too. TechCrunch noted that Square already uses voice AI for ordering, and said an on-device voice model is a possible path to more business and consumer uses at lower cost.

Some claims are still unproven. There are no independent tests yet of Dessa's quality, speed or safety. How much speech data was used to train it is not known.

What happens next

Block is asking outside experts to look at the model first. The company has issued a call for safety researchers to test Dessa, TechCrunch said.

When or whether Dessa will be released to the public, and on what terms, is not known. It is also not known whether any tags or controls steer its emotions, a point TechCrunch said remains unclear.

FAQ

What is Dessa? A voice model from Block's research lab that is trained fully on human speech.

What does full-duplex mean? The model can talk and listen at the same time.

Can it run on a laptop? Block says it runs on a MacBook M4 Max or better.

Can I use it now? No. There is no public access at the moment.

Who leads the project? Ragavan Thurairatnam leads Dessa, and the lab reports directly to Jack Dorsey.