Introduction

Dessa is a new speech model from Block, the company behind Square and Cash App. Block's small in-house AI lab

with a launch page full of live recordings, and Jack Dorsey
an hour later. It's a full-duplex speech-to-speech model: audio goes in, audio comes out, and it listens and talks at the same time, the way people do.

The headline claim is how it was trained. Block says Dessa learned only from recorded human conversation, with "no text, no pretrained models" underneath. This is a research preview, not a product. There are no weights, no API and no app yet. Block is giving early access to a small group of AI safety researchers first, and says an app is coming.

What Block Showed

Every clip on the launch page is a live recording of a person talking to Dessa, with subtitles added afterward. Block says nobody steers it, and Dessa picks what to say, when to say it, and in what voice.

  • Laughing mid-sentence: it laughs the moment something is funny, instead of waiting for you to finish.
  • Volume on request: ask it to speak up and it gets louder. Block says "nobody taught it that."
  • Finishing your thought: when a speaker blanks on a movie title, Dessa fills it in.
  • Emotional reads: in an anniversary clip it gasps "oh my god" under its breath and answers with what sounds like teary gratitude.
  • Backchannels: it drops in "yeah!", "got it!" and "ooh" while you're still talking, to show it's following.
  • AI-to-AI: one demo has Dessa chatting with GPT-Live-1, a third-party voice model.

The Specs Behind It

Block hasn't published a paper, parameter count or benchmark table. Here's what the launch page and the lab lead actually state:

  • Modality: true speech-to-speech, full-duplex. No speech-to-text or text-to-speech step in the middle.
  • Training data: human conversation only. Block says it learned language from "hundreds of times fewer words" than an LLM needs.
  • Training compute: a single NVIDIA GB200 NVL72 rack.
  • Speed: inference runs faster than real time.
  • Local hardware: runs on a MacBook Pro M4 Max or better. Lab lead Ragavan Thurairatnam
    it runs on consumer GPUs.
  • Team: a small lab founded by former members of Dessa, the Toronto deep learning startup, which has built generative speech models for over eight years and reports directly to Dorsey.

Why It Matters for Voice Work

Most voice assistants still turn your speech into text, think in text, then read the answer back. That middle step throws away tone, timing, sighs and laughter, which is why so many voice agents feel stilted and wait politely for silence before they answer. Speech-to-speech models like ElevenLabs' Eleven Realtime v1 and Grok Voice 3 already skip that step at runtime, but they still sit on top of text-trained language models.

Dessa goes further by never seeing text at all. If that holds up, it's a different starting point for anyone building characters, game NPCs or interactive audio: a model whose sense of timing and emotion came from conversation itself. Block also frames it as a safety question, since an AI that can't hear pain or frustration in a voice will keep talking over it.

Keep the limits in mind. All we have are hand-picked demo recordings. One early listener on X noted that it mixes up names, which fits a model that learned every fact from raw sound.

How to Get Access

You can't run Dessa today. Here's where things stand:

  • Listen now: the demo recordings, including two synchronized speaker tracks per clip, are on block.xyz/dessa. Block says more will be posted on X.
  • Researchers: Block is
    to probe the model. The page lists a contact address for the lab lead.
  • Everyone else: a team member says an "app coming soon." No date, price or license has been announced, and there's no word on open weights.

Conclusion

Dessa is the most interesting voice-AI idea of the week: a conversation model that learned to talk the way babies do, by listening, and that's small enough to run on a laptop. It's still a preview with no paper, no weights and no public app, so treat the demos as a promise rather than a benchmark. The things to watch are the app launch, whether Block publishes a technical report, and whether the weights ever leave the lab.

—Anabel ♡