People have long treated spoken language as the ordinary channel for exchanging facts, intentions, and feelings, yet most systems that process voice still convert what they hear into written text before deciding how to answer.
That conversion step discards timing, volume shifts, overlapping speech, and the small sounds that signal attention or hesitation.
Jack Dorsey, the co-founder of Twitter, before it became X after Elon Musk acquried, co-founded Block.
And now, his company has released a model named 'Dessa,'
The model, developed by its own internal research lab, is presented as a full-duplex speech-to-speech system.
Audio enters and audio leaves; no transcription layer sits in between. The lab states that the model was trained from scratch on recordings of human conversation alone, without text corpora or any pretrained models. Block describes it as, to the best of its knowledge, the first system able to sustain realistic live conversation under that constraint.
Training took place on a single Nvidia GB200 NVL72 system.
The lab notes that language was acquired from a quantity of words hundreds of times smaller than the volumes typically used for large language models, while the model also had to capture accents, breathing, sarcasm, laughter, and the moments when one speaker may cut in or offer a brief acknowledgment.
The stated reasons for the work include simple curiosity about whether a model could learn speech the way infants do, without ever seeing written words, along with a desire to retain the emotional and temporal information that text pipelines discard, and a safety concern that systems unable to register distress in a voice might continue regardless.
The project page hosts live recordings of people speaking with the model, subtitles added afterward.
In the clips the model laughs mid-sentence when something is funny, finishes a forgotten movie title, and produces responses the page characterizes as empathetic. One recording shows it in conversation with OpenAI's GPT-Live-1.
Observers have remarked that some exchanges sound largely natural while others can strike listeners as awkward or uncanny.
No parameter count, exact size of the training audio, or formal benchmark results have been published.
Early access is limited to a small group of AI safety researchers who apply with institutional addresses; a public waitlist is open and broader distribution is described as forthcoming.
The site itself lists open questions about how efficiently speech-only models acquire language and facts relative to text-trained systems, how readily they can learn additional languages, and what kinds of in-context learning they can perform.
But one of the best things about it is that, inference is said to run faster than real time, and the model is reported to execute locally on a MacBook Pro equipped with an M4 Max chip or comparable hardware.
The approach therefore sits at some distance from the dominant text-centered pipelines used in most current voice interfaces.
Dessa was originally a small research group, formed in Toronto. It was acquired by Block in 2020, and now the team reports directly to Dorsey.
