Block unveils Dessa, a voice model that speaks before thinking

Block introduced Dessa, a voice AI model trained solely on human conversation. The company describes it as a real-time speech-to-speech system designed to better capture emotion and test security challenges.

On October 23, 2026, Block unveiled Dessa, a voice AI model designed to converse in real time from simple human conversations. Jack Dorsey’s company says it trained the model without text or preexisting models and aims both to show what voice can still bring to AI and to open a research path on security. The idea is simple, almost startling: make a machine speak without having it read a line.

Une IA qui écoute et répond sans passer par le texte

Block describes Dessa as a speech-to-speech model, meaning audio input, audio output, without detouring through text transcription. The system is also full-duplex: it can listen and speak simultaneously, instead of waiting for a sentence to end before responding. According to the presentation page, it has learned to interrupt at the right moment, to give listening cues, and to handle laughter, sighs, sarcasm, or emotion in the voice.

The company states that Dessa was trained “from scratch on nothing but human conversation,” without text or prior models. This approach differs from classic voice assistants, which often convert speech to text before processing the request. Here, Block wants to preserve what text erases: rhythm, hesitations, and tone variations.

On its page, the company goes further and presents Dessa as a model capable of running locally on a MacBook Pro M4 Max or more powerful. Ragavan Thurairatnam, head of Dessa, speaks on X of “a GPT3 moment for speech.” A phrase that says a lot about the ambition, even if the comparison with GPT-3 remains, at this stage, more a metaphor than a scientific measure [to be verified].

Des démonstrations qui veulent prouver l’effet de présence

Block has published several recordings presented as real conversations with Dessa. In one, the model first mistakes the identity of its interlocutor, then smoothly resumes the discussion. In another, it improvises on Christopher Nolan before drifting to Bryan Singer’s Superman films. None of this resembles a perfectly smooth demo. It’s quite the opposite, and that is probably the point.

The company emphasizes that the model chooses what to say, when to speak, and which voice to adopt by itself. Subtitles were added afterward, Block specifies. In one video, Dessa bursts out laughing mid-sentence; in another, it reacts to a surprise announcement with what the page describes as “teary” gratitude. It’s not telepathy, just a more credible vocal synchronization effort than usual.

However, these demonstrations remain excerpts selected by the company. They mainly show what the product aims to do: reduce the impression of latency, mechanicalness, and emotional distance that still clings to many voice assistants.

Pourquoi Block mise sur la voix, et pas seulement pour faire joli

Block gives three reasons for this work: curiosity, better understanding between humans and AI, and security. The security argument is central. The company believes current voice systems “flatten what we say into text” and miss essential signals like pain, annoyance, or fear. In its typical scenario, an assistant would keep talking without understanding that a user is in distress. Not ideal, even for a polite machine.

In this context, Dessa also serves as a testbed. Block announces early access to a small group of AI security researchers, via an application targeting academic or lab profiles. The company also opens a public waiting list and promises more excerpts in the coming days.

The project comes from a small internal research team at Block, originating from the former Toronto startup Dessa. Thurairatnam says he founded and sold that company before joining Block. The group includes several researchers and former interns and reports directly to Jack Dorsey. In other words, this is not just another demo from an anonymous lab but work that Block explicitly ties to its in-house AI strategy.

Ce que révèle ce prototype sur la prochaine bataille des assistants

Dessa not only says something about Block. The model also points to a broader battle: who will make voice a useful, natural, and reliable interface, rather than just a chatbot skin. Text models have already set their pace; voice models are still searching for the right grammar. And timing matters, as voice is becoming central again in hands-free use, embedded assistance, and more sensitive interactions.

Block is making a bold technical bet here: learning language from speech alone, with far fewer words than large text models, and running it in real time on a consumer machine. It remains to be seen what the limited access promised to researchers will yield, then the wider rollout. For now, Dessa looks less like a finished product and more like a demonstration of direction. But a direction where AI speaks, listens, and interrupts less, which, let’s admit, won’t be the case for everyone.

By the numbers

  • October 23, 2026 — date of Block’s announcement.
  • 1 speech-to-speech model presented: Dessa.
  • 0 text used in training, according to Block.
  • 1 small group of security researchers receives early access.
  • 1 MacBook Pro M4 Max or better for announced local execution.

To read