Dessa, Block's voice that learned to speak without reading

On October 23, Block, the company behind Square and Cash App led by Jack Dorsey, introduced Dessa as a research preview. It is a speech-to-speech model: according to the company's official page, it takes audio in and gives audio back without going through a transcript, and it keeps listening while it talks. Block calls it “the most authentic full-duplex speech model ever built” and, with the qualifier “to our knowledge”, presents it as the first model of its kind. What sets it apart, again according to the company, is how it was trained: from scratch, only on recorded human conversations, with no text and no pretrained models.

That choice turns the usual approach on its head. Many voice assistants convert speech into text, or start from a text-based language model and layer audio on top, and in that step tone, pauses and emotion tend to get lost. This is exactly the point Block presses, noting that today's most powerful AI systems “can't hear how you feel”. The company says Dessa has learned when to interrupt and when to drop in short listening cues, runs faster than real time, and needed “hundreds of times fewer words” than language models. On infrastructure, Block reports a single NVIDIA GB200 NVL72 system for training, and local inference on a MacBook Pro M4 Max or better. The usual points of comparison in this field are Kyutai's open-weight Moshi and NVIDIA's PersonaPlex, which is built on Moshi.

The scope of what can be verified, though, is narrow, and that needs saying plainly. Every performance claim comes from Block, backed by demo recordings the company itself chose and published: conversations about films, a guessing game, dialogues with GPT-Live-1, a third-party model whose developer, Block notes, it is neither affiliated with nor endorsed by. Crypto Briefing points out that some of these demos also show limits, such as speakers' identities getting swapped or abrupt changes of topic. According to Implicator and Crypto Briefing, there is no parameter count, no weights, no pricing, no API and no standard benchmarks, and no independent evaluations exist yet. We don't know how much data was used, which languages are supported, when access might widen, or whether the weights will ever be released. Without the figure on training data, even the “hundreds of times fewer words” comparison can't be checked. For now there is a waitlist, and early access is limited to a small group of AI safety researchers with institutional email addresses. Behind the project is a lab Block itself describes as very small, founded by former members of Dessa, the Toronto deep learning startup that shares the model's name, and, according to the company, reporting directly to Dorsey.

What draws me in is the idea of a machine that reaches language through the ear rather than the page, because it shifts the question from what is said to how it is said. But a voice that knows when to stay quiet also has to be heard by people who didn't build it, and as long as the only recordings available are the ones Block picked, what we mostly know about Dessa is the story the company tells. I'm waiting for the moment someone else gets to talk to it and hear what it says back.

— Olya