Block unveils Dessa, a voice model trained only on recorded human conversation

Block, the company behind Square and Cash App, has shown Dessa, an experimental speech-to-speech model from a small research lab inside the company. Block says Dessa was trained from scratch on recordings of human conversation alone, with no text and no pretrained models. The company also says the model runs locally on a MacBook Pro with an M4 Max chip or better.

Most voice assistants turn speech into text, have a language model process the text, and then turn the reply back into audio. Block argues that this pipeline throws away tone, hesitation and emotion. If a model can hear and produce those signals itself, and run on one laptop, it changes the cost and the feel of voice interfaces in customer service, ordering and personal assistants.

What Block says Dessa is

Dessa is a full-duplex model: it listens and speaks at the same time, so it can be interrupted, give small acknowledgements and react before the other person has finished. TechCrunch notes that OpenAI released a full-duplex model, GPT-Live, earlier this year.

On its announcement page, Block calls Dessa "the most authentic full-duplex speech model ever built" and writes: "We trained Dessa from scratch on nothing but human conversation." The lab says the model learned language from "hundreds of times fewer words" than large language models, and that one spoken word takes "thousands of times more bytes" than the same word as text. These figures come from Block and have not been independently checked.

Block describes the problem it is trying to solve in plain terms. "Interacting with current voice AI can just plain suck. It turns your voice into text and throws away an incredible amount of information text alone can't carry," the company wrote, according to TechCrunch. "The most powerful AIs in the world can't hear how you feel."

TechCrunch points out that products such as ElevenLabs v4 and Fish Audio Drama 3 use voice tags to steer emotion. Block says emotion is built into Dessa. TechCrunch adds that it is not clear whether some form of tagging steers it behind the scenes.

Training, hardware and the team

Block says it trained Dessa on "only one NVIDIA GB200 NVL72." Notebookcheck describes this as a single system that holds 72 Blackwell GPUs. For inference, Block says the model runs on a MacBook Pro M4 Max or better, and Notebookcheck reports that it runs faster than real time.

The page describes the group as "a (very) small AI research lab at Block." Former members of Dessa, a Toronto deep learning startup, founded it. Ragavan Thurairatnam leads the lab, which reports to Jack Dorsey. The page lists seven researchers and several current and former interns. The lab says it has built deep learning models since 2012 and generative speech models for more than eight years.

Block has not published a parameter count, model weights, benchmark results or pricing.

What the demos show, and what they do not

Block's researchers posted recorded conversations in which Dessa uses different voices, speeds and moods. In one, the model says it is tired from work and sounds tired. In another, it sounds close to breaking down as it talks about the person making it feel comfortable. A third shows Cara Ji, an intern on the team, playing a guessing game with the model. Block also shows Dessa talking with OpenAI's GPT-Live-1 and says that developer "is not affiliated with Block and has not reviewed or endorsed Dessa."

All of these are first-party recordings. Block has not said how the clips were chosen or how often conversations go this well.

The model is not publicly available. Block has opened a waitlist and says "launch links are coming soon." A small group of AI safety researchers will get early access to find risks the team has not anticipated. "A model that learned to talk from sound alone is new territory," the page says. TechCrunch suggests the testing is likely to cover cases where a voice agent could deceive the person on the other end of a call.

What it means for work

Block already uses voice AI in Square to take orders. TechCrunch notes that a voice model the company owns and can run on a device could power more business and consumer services at lower cost. The tasks most exposed are spoken, routine and high-volume: taking orders, answering store and support calls, booking appointments, and first-line customer service. Those jobs are done today by cashiers, call centre agents and receptionists. A model that sounds tired, amused or sympathetic also makes it harder for a caller to tell whether a person is on the line, which is why the safety review and disclosure rules matter as much as the voice itself. What remains open is how reliable Dessa is outside curated demos, and whether a model trained only on conversation, without a text model for knowledge, can handle the factual parts of these jobs. It is also unclear when, and in what form, it will reach products.