Block Says Its New Voice Model Learned to Talk Without Ever Reading a Word

Block has announced Dessa, a voice model it says it trained from scratch on recorded human conversation alone, with no text and no pretrained model. The announcement went up Oct. 23, 2026. Block, the payments company that owns Square and Cash App, opened a waitlist rather than public access, and nobody outside the company has tested the model.

Block says Dessa takes audio in and puts audio out with no transcription step in between, and that it is full duplex: both sides can speak at once, the way people do in ordinary conversation. The company's case is that most voice assistants convert speech to text first, discarding tone, timing and hesitation. It says a model trained on sound can hear emotion, sarcasm, laughter and a sigh directly.

The company says the model generates speech faster than real time and runs locally on a MacBook Pro M4 Max or better. It says it trained Dessa on a single NVIDIA GB200 NVL72 system, and that the model learned language from hundreds of times fewer words than a text-trained large language model needs. Block publishes no benchmark scores and has not said how much recorded speech it used.

The evidence on Block's announcement page is a set of recordings that it describes as live, unsteered conversations with a person, subtitles added afterward. In one, Dessa talks to GPT-Live-1, a third-party model; Block says its developer is not affiliated with Block and has not reviewed or endorsed Dessa. No independent evaluation of the model exists. Block says it is giving a small group of AI safety researchers early access to find risks it has not anticipated; everyone else can join the waitlist.

Block says that, as far as it knows, Dessa is the first AI, trained only from human speech, that a person can hold a realistic live conversation with. The lab is a small research group inside Block, led by Ragavan Thurairatnam and founded by former members of Dessa, a Toronto deep learning startup of the same name.