Dessa: Block's Unreleased Full-Duplex Voice AI, Explained
Dessa is Block's unreleased speech-to-speech model trained on audio alone. Here's what it claims and how natural it actually sounds.
What is Dessa, and who built it?
Dessa is an unreleased voice AI model from a small research lab inside Block, the fintech company run by Jack Dorsey that owns Square and Cash App. The lab is led by a researcher named Ragavan, and according to the team, it was founded by former members of a Toronto deep learning startup that also went by the name Dessa. The team says it has worked on deep learning since 2012 and on generative speech models specifically for more than eight years. Dessa has not been released publicly. What exists right now is a set of demo audio clips and blog claims, not a model anyone outside Block can test.
TL;DR
- Dessa is a speech-to-speech model that Block says is trained from scratch on human conversation audio, with no text data and no pretrained language model underneath it.
- Full duplex operation means it can listen and talk at the same time, letting it interject with “yeah,” laugh mid-sentence, or cut in at a natural pause instead of waiting silently for its turn.
- The architecture is not public. Block has shared a general conceptual diagram but no detailed technical paper, so independent verification of how it works is not currently possible.
- Hardware claims include local, real-time operation on a MacBook Pro with an M4 Max chip or better, with training reportedly done on a single Nvidia GB200 NVL72 system, both according to Block’s own blog.
- Block claims major data efficiency, saying Dessa learned language from hundreds of times fewer words than a typical large language model, while acknowledging that spoken audio is inherently harder to model than text.
- The demo audio is a mixed bag. Some exchanges sound close to natural conversational flow, including a genuinely convincing laugh, while other segments, especially a male voice in several clips, sound noticeably synthetic or “plasticky.”
- None of Block’s performance numbers have been independently verified because the model is not yet available for hands-on testing.
Remy doesn't build the plumbing. It inherits it.
Other agents wire up auth, databases, models, and integrations from scratch every time you ask them to build something.
Remy ships with all of it from MindStudio — so every cycle goes into the app you actually want.
How is Dessa different from typical voice assistants?
Most voice assistants today, including the ones layered on top of popular chatbots, work as a three-stage pipeline. Your spoken audio gets transcribed into text by a speech recognition system. A language model reads that text and generates a text reply. A separate text-to-speech system reads that reply out loud. This is sometimes called a cascaded or pipeline approach, and it is how a large share of voice products are built today because each stage can be swapped out and improved independently.
The problem, according to Block’s team, is that converting speech to text throws away information that never makes it back. Tone of voice, timing, pauses, a sigh, sarcasm, a laugh landing mid-thought: none of that survives the trip through plain text. The language model never “hears” any of it, so it cannot react to it, and the text-to-speech layer has to reconstruct emotional nuance from scratch with no real signal to work from.
Dessa’s pitch is that it skips the text step entirely. It is described as a speech-to-speech model trained directly on audio of human conversation, with no transcripts and no underlying pretrained language model doing the heavy lifting. That matters because it means the system is, in theory, learning the relationship between sound and meaning directly rather than through a text intermediary.
What does “full duplex” actually mean here?
Full duplex means the model can listen and speak at the same time, the way two people do in an actual conversation. Most voice AI today is half duplex: it waits for you to stop talking, processes what you said, then responds. That turn-taking gap is a big part of why talking to AI assistants can feel stilted even when the words themselves sound natural.
In the demo clips, this shows up as small interjections, the model saying “yeah” or something similar while the other person is still talking, laughing at a joke mid-sentence rather than after a pause, or jumping in at what sounds like a natural conversational beat instead of waiting for dead air. Those are the moments that are hardest to fake convincingly, because they require the model to be processing incoming audio and generating outgoing audio simultaneously rather than in neat alternating turns.
What does Block actually claim, and what can’t be verified yet?
Block’s public claims, drawn from their blog and social posts, break down into a few categories:
Training data: Dessa is trained on human conversational audio alone, without text or a pretrained language model underneath. The team argues speech is a harder modality to learn from than text because a single spoken word carries far more raw data than its written form, and the same sentence spoken twice never sounds identical, so the model has to generalize across far more variation per unit of meaning.
Efficiency: despite that difficulty, Block says Dessa learned language from hundreds of times fewer words than a typical large language model needs.
Hardware: the blog states Dessa runs faster than real time and can run locally on a MacBook Pro with an M4 Max chip or better. Training reportedly happened on a single Nvidia GB200 NVL72 system.
None of these figures come with a published technical paper, benchmark suite, or third-party evaluation. They are Block’s own numbers, and the only way to confirm them is to get hands-on access to the model, which is not currently possible since Dessa has not been released. Until it is, these claims should be treated as a lab’s self-reported description of its own work, not as something independently confirmed.
Does the demo audio actually sound natural?
Mixed. Across the published samples, the results vary noticeably by clip and by voice. One early exchange, where the model mistakes a speaker for someone named “Ronaldo” and then corrects itself conversationally, flows fairly smoothly and captures some of the back-and-forth rhythm of real talk. A later clip where one speaker laughs about “winter coming” produces a laugh that lands naturally inside the sentence rather than after it, which is a genuinely hard effect to fake and one of the more convincing moments across the samples.
But several other clips expose rougher edges. A male voice responding with “I’m doing well” to “Hey, how you doing?” comes across as stiff and synthetic, what gets described as “plasticky” delivery, meaning the cadence and tone don’t track how a person would actually inflect that phrase. In another sample, the model’s female voice says “so blown away” and “it’s pretty much like human level” with a flatness that undercuts the enthusiasm of the words themselves. A separate test, where a speaker deliberately asks the model to “cough” to help it hear better, actually plays out convincingly, with the model adjusting and asking the person to repeat themselves clearly.
The overall picture is a model that nails some of the hardest parts of natural conversation, like mid-sentence laughter and timing, while still stumbling on baseline vocal naturalness in plainer exchanges. That inconsistency is worth keeping in mind against the framing on social media that this represents a huge leap for voice AI. Voice models have already improved substantially over the past few years, and judged purely on the demo clips, Dessa sits somewhere in that existing trajectory rather than clearly ahead of it.
Is Dessa worth paying attention to?
It depends on what you’re looking for. If you care about the underlying approach, speech-to-speech modeling trained on raw audio without a text bottleneck, Dessa is a notable data point, because full duplex, audio-native systems remain rare in the field and most production voice products still rely on the cascaded pipeline. If you care about whether it currently sounds better than alternatives like ElevenLabs or other voice models already on the market, the honest answer based on the available samples is: not clearly, and not consistently. Some clips are strong. Others are not. Nothing about Block’s efficiency, hardware, or training claims can be independently checked until the model is actually released.
Frequently Asked Questions
What is Dessa AI?
Other agents start typing. Remy starts asking.
Scoping, trade-offs, edge cases — the real work. Before a line of code.
Dessa is an unreleased speech-to-speech voice model developed by a small research lab at Block, the company led by Jack Dorsey. It is designed to process and generate spoken audio directly, without converting speech to text in between.
Is Dessa available to the public?
No. As of the source material for this article, Dessa has not been released. Only demo audio samples and blog claims from Block are publicly available, with no confirmed release date or access method.
How is Dessa different from models like ElevenLabs?
Most commercial voice tools, including ElevenLabs, generally focus on text-to-speech or work within a pipeline where speech is transcribed to text first. Dessa claims to be trained end-to-end on audio alone and to operate full duplex, meaning it can listen and speak simultaneously rather than strictly taking turns.
What hardware does Dessa claim to run on?
Block’s blog states Dessa can run faster than real time locally on a MacBook Pro with an M4 Max chip or better, and that it was trained on a single Nvidia GB200 NVL72 system. These figures come directly from Block and have not been independently verified.
Does the demo audio actually sound human?
It varies by clip. Some moments, like natural mid-sentence laughter, sound convincing. Other segments, particularly certain male voice responses, sound stiff or synthetic. The quality is inconsistent across the available samples rather than uniformly impressive.