Block’s speech-only, full-duplex research model. It learns from human conversation audio and is designed to listen while speaking.
Research preview · Download coming soon · No public API announced
Independent guide · Source checked October 26, 2026 · Release notes →
Timing. Tone. The spaces in between.
From quick reactions to longer exchanges, explore the timing, tone and conversational rhythm of Dessa. Play all six featured recordings here, with no account or microphone needed.
A conversation about the release, with attention to laughter, volume and changes in delivery.
This recording is temporarily unavailable. Please try again later.
Listen for how a response begins and ends. Does laughter fit the exchange? Are changes in tone connected to what the other speaker says?
A casual pop-culture exchange for exploring the rhythm of back-and-forth speech.
This recording is temporarily unavailable. Please try again later.
Pay attention to sentence completion, pauses and overlapping speech. A natural-sounding exchange does not by itself establish factual accuracy.
An expressive reaction to a personal surprise, illustrating how much meaning can travel through a voice.
This recording is temporarily unavailable. Please try again later.
Compare pitch, emphasis and pacing with the other recordings. Emotional-sounding speech is generated behavior, not evidence that the model experiences emotion.
A longer exchange that gives you more room to listen for continuity, pacing and response style.
This recording is temporarily unavailable. Please try again later.
Follow how each reply connects to earlier remarks. A selected conversation cannot establish dependable advice, safety or performance across users.
Coffee, a shared joke and a conversation about stress. Explore how vocal delivery shifts with the exchange.
This recording is temporarily unavailable. Please try again later.
Jump to sentence completion and the joke, then compare the longer discussion about stress. These are selected examples of expressive delivery.
A conversation between two AI-generated voices, with moments of interruption and short acknowledgments.
Dessa and GPT-Live-1 are AI-generated voices. The noisy recording also includes a human. No endorsement by the third-party model developer is implied.
This recording is temporarily unavailable. Please try again later.
Listen for how the systems take turns and recover from interruption. The recording does not imply any endorsement by the other model’s developer.
This recording is temporarily unavailable. Please try again later.
This recording is temporarily unavailable. Please try again later.
This recording is temporarily unavailable. Please try again later.
This recording is temporarily unavailable. Please try again later.
This recording is temporarily unavailable. Please try again later.
Dessa and GPT-Live-1 are AI-generated voices. The noisy recording also includes a human. No endorsement by the third-party model developer is implied.
This recording is temporarily unavailable. Please try again later.
This recording is temporarily unavailable. Please try again later.
This recording is temporarily unavailable. Please try again later.
This recording is temporarily unavailable. Please try again later.
This recording is temporarily unavailable. Please try again later.
This recording is temporarily unavailable. Please try again later.
This recording is temporarily unavailable. Please try again later.
This recording is temporarily unavailable. Please try again later.
This recording is temporarily unavailable. Please try again later.
This recording is temporarily unavailable. Please try again later.
Selected recorded demonstrations, not live model access or independent benchmarks. Playback starts only when you press play.
Read the full listening guide →A laugh mid-sentence. A change in pitch. Knowing when to jump in.
Dessa is Block’s speech-to-speech research model. Block describes this Dessa voice AI as full-duplex and trained from scratch on human conversation audio, without text or pretrained models. It is designed to listen and speak at the same time.
Those are the developer’s claims. The public evidence is a collection of selected conversation recordings; we have not independently tested the model.
Understand the approach →The developer reports faster-than-real-time inference on a MacBook Pro M4 Max or better. Model size, a full compatibility matrix and standardized performance benchmarks remain unpublished in this source snapshot.
Explore the training approach, capabilities & research questions →See the difference between strict turn-taking and overlapping conversation.
Illustrative diagram only. No audio or live AI.
Overlapping windows illustrate simultaneous listening and speaking.
Clear answers to the questions
behind the search.
A listening guide to the official recordings.
Read the guide →FIELDNOTE 02Download, release, and access status.
Read the guide →FIELDNOTE 03What is known about API availability.
Read the guide →FIELDNOTE 04The M4 Max claim, and what is still unknown.
Read the guide →Listen to 21 complete recordings, explore how simultaneous listening and speaking works, and check what is known about model access. This guide explains the research and does not run the model or manage its waitlist.
How Dessa works →Dessa comes from a small research lab at Block led by Ragavan Thurairatnam. The team says it was founded by former members of the Toronto startup Dessa and reports directly to Jack Dorsey.
The team describes experience building deep-learning models since 2012 and more than eight years of work on generative speech. Read the team background and our editorial method below.
Team context & editorial method →Dessa is a speech-to-speech research model developed by a small lab at Block. Block says it learned from human conversation audio without text or pretrained models.
Block provides conversation recordings. Public model access is not available through this website. The official page describes limited early access for safety researchers.
The official download control says coming soon as of October 26, 2026. See our download guide for the distinction between demos and model availability.
The official page we checked does not announce public pricing or an open-source license. We do not sell access.
No public API documentation is provided on the official page we checked.
A small research lab at Block led by Ragavan Thurairatnam, founded by former members of the Toronto startup Dessa. The team says it reports directly to Jack Dorsey.