An independent guide to Block's voice model

Dessa AI. Beyond the words.

A laugh. A pause. A change in tone. Meet Block's speech-to-speech research model, trained on human conversation — and explore what it can actually do today.

Research preview · Public access not confirmed

Sculptural sound waves in forest green and terracotta facing one another.
Two voices, one conversation. Concept illustration.
Developed by
Block
Input → output
Speech → speech
Conversation
Full duplex
Current stage
Research preview
Source: BlockChecked October 25, 2026

Get to know Dessa AI

More than what you say. How you say it.

Block is exploring what a model can learn from the sound of human conversation. Here is the idea behind the research, and where the evidence ends.

01

Learning from conversation

Block says Dessa was trained from scratch on human speech, without text or pretrained models. That is a training claim, not an independently audited result.

02

Listening while speaking

Full duplex means listening and speaking can overlap. The demos explore interruptions, short acknowledgments and conversational timing.

03

A voice with expression

Block highlights laughter, pauses and changes in tone in its selected recordings. Sounding expressive does not establish emotional understanding or reliability.

Why is Dessa AI getting attention now?

Dessa AI was introduced in a research-team post on October 23, 2026 at 16:38 UTC, followed by Jack Dorsey's post and an invitation to safety researchers. The speech-only training claim and expressive recordings explain what people are discussing. These dated events establish the research launch; they do not establish broad public access or measure how popular the model is.

Read the research announcement

The story behind Dessa AI

A company name, an earlier speech project and a new conversation model are connected by people, not a proven shared architecture. Dates below identify announcements, not permission to use the new model.

  1. RealTalk: speech from text

    The original Dessa company introduced RealTalk as text-to-speech research. That project is separate from the new model's speech-conversation task.

    Original RealTalk introduction
  2. Square acquires the company

    Square announced its acquisition of the Toronto machine-learning company Dessa. This is company history, not a model release date.

    Square acquisition announcement
  3. Square announces the Block name

    The company announced its new Block name while retaining Square for its seller business. This explains the two publisher names in the history.

    Company name announcement
  4. Dessa AI research announcement

    The team introduced the new full-duplex speech model at 16:38 UTC. Block says the lab includes former Dessa members; it has not documented the new model as a rename of RealTalk.

    Research team’s original post
Understand the technology

A closer listen

The interesting part is in the timing.

Hear the official recordings with a little context. A chosen example can show a behavior; it cannot tell us how often a model gets it right.

Official recordings · Hosted by Block
Explore the Dessa AI demos

One voice

Another voice

An illustration of overlapping conversation, not measured model latency.

A few things worth knowing.

The short answers before you explore Dessa AI further.

Follow the evidence

Know what changes. Know what comes next.

Public access, API documentation and model downloads are separate milestones. Our update log keeps the changes and their sources together.

Cookie preferences

Necessary cookies support sign-in, payments and security. Choose whether to allow analytics, ads and third-party chat. You can change your choice in the footer. Privacy policy