TL;DR
- Speech Training: Block unveiled Dessa, a voice AI it says learned to converse from human speech without text or pretrained models.
- Conversation Limits: Selected demos combine expressive voices with mistakes, including continued guesses after the model is told it has won.
- Existing Alternatives: Kyutai’s Moshi and NVIDIA’s PersonaPlex already listen and speak simultaneously, using language models trained on text.
- Research Access: Block is inviting a small group of AI safety researchers; Dessa is not publicly available.
Block has unveiled Dessa, a voice AI model it says learned from human speech alone to hold conversations that include interruptions, laughter and changes in tone. The company is inviting AI safety researchers to try the model through a small early-access program.
The new AI model comes from a research lab inside Block, the company behind Square and Cash App. Ragavan Thurairatnam leads the lab, which reports directly to Block cofounder Jack Dorsey. Its model and the lab share the Dessa name.
We made the first AI model you can have a realistic conversation with that was trained entirely on speech!
Dessa is a full-duplex speech-to-speech model that you can run on consumer GPUs.
This is a GPT3 moment for speech. pic.twitter.com/XVD8sSyJn8
— Ragavan Thurairatnam (@RagavanThurair) October 23, 2026
Learning the Rhythm of Conversation
Dessa takes in audio and produces audio. Block says it trained the model solely on human conversations, without text or a pretrained model supplying an existing language foundation.
Dessa is full-duplex, meaning it can keep listening while it speaks. That lets a conversation include overlapping speech and brief acknowledgments before the other person finishes. The lab says Dessa chooses its own words, timing, voice and personality in the recordings it published.
The company describes laughter midway through a person’s sentence and a voice that grows louder when asked to speak up. Those changes in delivery are part of what the team wants Dessa to learn directly from sound, alongside the words themselves.
Meet Dessa, our AI model trained from scratch on nothing but speech.
We believe speech is the natural interface for technology. So we built Dessa.
And we’re inviting AI safety researchers to explore Dessa.
Learn more: https://t.co/6wA8KhWdjn pic.twitter.com/vXM43uUFzx
— Ragavan Thurairatnam (@RagavanThurair) October 23, 2026
Natural Speech, Uneven Answers
Thurairatnam also claims Dessa acquired world knowledge and learned to reason from audio. In a guessing-game recording he shared, a Block intern asks Dessa to identify everyday objects.
Dessa developed world knowledge and learned to reason purely from audio
Here’s our intern @carajzp playing a game with it pic.twitter.com/dTvIe8FfVQ
— Ragavan Thurairatnam (@RagavanThurair) October 23, 2026
Dessa also might mistake its conversation partner for someone named Ronaldo as reported by The Verge. In another clip, a discussion of Christopher Nolan’s films veers into Bryan Singer’s Superman films.
What Moshi and PersonaPlex Already Do
Simultaneous listening and speaking already exist in Kyutai’s Moshi. It models separate audio streams for the person and the model, while predicting text corresponding to its own speech. A text language model supplies its underlying foundation, helping it produce spoken responses.
NVIDIA’s PersonaPlex builds on Moshi’s architecture. It lets a developer specify a role through text and voice characteristics through an audio sample. NVIDIA provides code and model weights under separate licenses, giving developers a released implementation alongside those controls.
Where Moshi and PersonaPlex combine spoken interaction with text-based foundations, Block is asking whether speech alone can supply both conversational behavior and knowledge.
Laptop Inference and Rack-Scale Training
Block says a trained Dessa model can run locally on a MacBook Pro M4 Max or better, producing audio faster than real time. Moshi’s maintainers report testing its on-device implementation on a MacBook Pro M3.
Dessa lab says it trained the model on one NVIDIA GB200 NVL72, a rack containing 72 GPUs. Block does not identify the conversation dataset or how permission to use it was obtained in the announcement.
Why the Lab Wants Outside Testing
The lab describes curiosity, better communication and safety as its reasons for the project. The team says it wanted to see whether a model could learn to talk as children do, before encountering written words.
Dessa lab imagines for example someone in serious pain whose distress goes unnoticed by a voice system that keeps going as a use case. Learning emotional cues from sound is intended to help a model respond to the person’s condition as well as their words.
How reliably Dessa can recognize that distress remains an open question. As of October 24, it is not publicly available. Block is inviting AI safety researchers to look for risks the team has not anticipated.