Block, the company led by Jack Dorsey, announced an AI model called Dessa on Friday that it says was trained “from scratch on nothing but human conversation.” The Verge reported that Block showed the model in recorded voice conversations that are mostly natural but occasionally awkward. The announcement matters because most voice assistants convert speech to text and back, while Dessa is described as a true speech-to-speech model, and because Block says it can run locally on a MacBook Pro.

What Block says Dessa is

According to Dessa’s website as quoted by The Verge, the model is a “true speech-to-speech model, and everything it knows came from human speech.” That is different from the common design in which a speech-recognition system transcribes audio, a language model writes a text reply and a speech synthesizer reads it aloud. Block’s team argues that such pipelines “flatten” what people say into text and “can’t keep up with human timing.”

Block says Dessa can run locally on a MacBook Pro with an M4 Max chip “or better.” The coverage does not state the model’s size, license or whether the weights will be released, and we have not seen a technical report.

The demonstrations

The Verge describes several examples published by the team. In one, Dessa plays a guessing game over voice with a speaker, though it initially mistakes the speaker for someone named “Ronaldo.” In another, a conversation about Christopher Nolan’s films drifts into an unexpected tangent about Bryan Singer’s Superman films. A third has a person telling Dessa about taking it on a surprise trip to Paris, after which, in the website’s description, the model answers with what sounds like teary gratitude. The site also includes a conversation between OpenAI’s GPT-Live-1 voice model and Dessa, in which Dessa comes across as more natural, according to The Verge.

These are selected clips released by the developer, not a benchmark. Without independent tests of latency, accuracy, safety behavior and failure rates, they show what the model can do at its best, not how it performs in general.

Why Block says it built it

The team lists three reasons: “curiosity (aka fun), better understanding between humans and AI, and safety.” Its explanation draws an analogy: babies learn to talk without seeing written words, and the team wanted to see whether a model could too. It argues that speech carries emotion and timing that text pipelines discard, and that a voice AI deaf to emotion is a safety problem, for example if someone speaks while in serious pain and the system carries on regardless.

Access

Block says it will give a small group of AI safety researchers early access to Dessa. There is no general release described in the reporting we reviewed, and no pricing or product plans.

Why it matters

Voice is a fast-moving area, with several labs shipping real-time speech models. A model trained only on speech, if the claim holds, would be a notable departure from approaches that reuse text-trained language models. The claim that it runs on a laptop-class chip would also matter for privacy and latency, since a local model does not need to send audio to a server; see our explainer on cloud versus local model trade-offs. Equally, “knows only what is in human speech” raises open questions about factual knowledge, which usually comes from large text corpora.

What to watch next

  • A technical report or model card describing the training data, size and evaluation.
  • Whether Block releases weights or an API beyond the safety-researcher preview.
  • Independent latency and quality comparisons against other speech-to-speech models.
  • How the model handles factual questions, given its speech-only training.
  • Findings from the safety researchers who get early access.

Key takeaways

  • Block announced Dessa, a speech-to-speech model it says was trained only on human conversation.
  • The company says it runs locally on a MacBook Pro M4 Max or better.
  • Access is limited to a small group of AI safety researchers for now.
  • Evidence so far is developer-selected demo clips; no benchmarks or paper were cited.

Frequently asked questions

What is a speech-to-speech model?

It takes audio in and produces audio out directly, instead of converting speech to text, generating a text reply and converting that to speech.

Can I try Dessa?

Not generally. Block says early access will go to a small group of AI safety researchers.

Is Dessa open source?

The coverage we reviewed does not say. No license or weights release was mentioned.