Block's Dessa Learned to Talk Without Reading a Word of Text

Block, the company behind Square and Cash App, has unveiled Dessa, a voice model that learned to hold a conversation only by listening to people talk. Announced on October 23, 2026, Dessa is a full-duplex speech-to-speech model trained from scratch on human conversation, with no text and no pretrained models. Block says it is, as far as it knows, the first AI trained only on human speech that can hold a realistic live conversation. Block announced it the same day from its @blocks account on X. Dessa is a research preview, not an open-weights release: Block hasn’t published the model or its weights, and for now it is offering early access only to a small group of AI safety researchers.

Audio in, audio out

Most voice assistants work as a pipeline. Speech is turned into text, a language model writes a reply, and a speech engine reads that reply out. Block’s argument is that the first step throws away much of what people say. Tone, sarcasm, laughter, hesitation and a frustrated sigh can all change what a sentence means, and none of it survives the trip through text.

Dessa skips the text step. It takes in audio and produces audio, and it is full-duplex, so it listens and speaks at the same time instead of taking strict turns. Because it learned from real conversations, Block says, it picks up when to cut in and when to offer short backchannels such as “yeah” or “got it” while the other person is still talking. Block says inference runs faster than real time and that the model runs locally on a MacBook Pro with an M4 Max chip or better.

The company’s recordings are meant to show behaviour no one programmed. Block says Dessa laughs in the middle of a sentence when something is funny, gets louder when asked to speak up, finishes a sentence when a person can’t recall a movie title, and reacts to a surprise with a whispered “oh my god”. Every clip is a live recording of a person talking to the model, with subtitles added afterwards. One clip is a conversation between Dessa and ChatGPT’s GPT-Live-1 voice model; Block notes that the developer of GPT-Live-1 is not affiliated with it and has not reviewed or endorsed Dessa.

Learning language from sound

Training on speech alone is much harder than training on text, and Block’s page explains why. The same sentence spoken twice produces different data each time, and a spoken word takes thousands of times more bytes than the same word written down. The model also has to learn accents, pauses, breathing, emphasis and timing alongside the words themselves.

Block says Dessa learned language from hundreds of times fewer words than large language models use. It says it also worked out how to train the model on a single NVIDIA GB200 NVL72 system. Block hasn’t published a parameter count, a training-data size or benchmark results. The page instead raises open research questions, including how efficiently a speech-only model learns facts and new languages compared with a text model, and what kinds of in-context learning it can do.

Why Block is calling in safety researchers

Block gives three reasons for building Dessa: curiosity, better understanding between people and AI, and safety. On the last point, it argues that voice AI that can’t hear emotion is a risk in itself, because an assistant may carry on cheerfully while someone on the line is in serious pain. A model that hears distress directly is, in Block’s framing, a step toward fixing that.

A model that sounds this human also raises the opposite concern: that it could be used to impersonate people or run convincing phone scams. Block hasn’t set out specific misuse safeguards on the announcement page. Instead it is asking outside researchers to find the risks it hasn’t anticipated before any wider release, describing a model that learned to talk from sound alone as “new territory”.

A small lab with a long history

Dessa comes from what Block calls a very small AI research lab, founded by former members of Dessa, the Toronto deep learning company. The team says it has built deep learning models since 2012 and generative speech models for more than eight years, and that it reports directly to Block co-founder and chairman Jack Dorsey. The project is led by Ragavan Thurairatnam. Block says more recordings will be shared on X, and researchers who want early access can contact the team directly.