Block's new Dessa voice model learned to talk just by listening

If you've spent any time talking to voice AI, you know that uncanny, read-from-a-script feeling. Even the most polished assistants usually rely on the same tired loop: transcribe your voice to text, feed that text into an LLM, and synthesize a voice back out. In that middle step, almost all human emotion gets discarded.

Jack Dorsey's Block wants to kill that pipeline. On October 24, the company's research lab unveiled Dessa, a native real-time voice model that learned to converse without using any text or traditional pre-training models at all. Much like a human baby, it learned vocabulary, grammar, and emotional nuance simply by listening to thousands of hours of human conversation.

Full-duplex banter that laughs, pauses, and cries

In a blog post, Block was blunt about the state of the industry, saying that interacting with current voice AI can just plain suck because throwing away vocal tone creates an unnatural and potentially dangerous barrier. While tools like ElevenLabs v4 or Fish Audio Drama 3 often rely on explicit voice tags to simulate mood swings, Block claims Dessa's emotional responsiveness is built directly into its architecture.

Dessa is a full-duplex model, meaning it talks and listens simultaneously, just like OpenAI's GPT-Live. You can cut in while it is speaking, and it catches conversational cues seamlessly. In shared demos, the bot sounded genuinely tired when discussing a long workday, giggled at jokes, and even sounded close to breaking down in tears when having an emotional chat. When a speaker was too quiet, Dessa casually asked them to move closer to the mic so it could hear better.

Remarkably, Block claims the entire model was trained on a single Nvidia GP200 GPU rack. It doesn't need massive data centers to run inference, either; it can run locally on consumer-grade hardware like a MacBook with an M4 Max chip. For now, public access remains shut. Block is only opening test access to AI safety researchers to examine risks before considering broader availability.

Why it matters

We think Dessa represents an exciting architectural pivot away from the text-first orthodoxy that has dominated generative AI for the past two years.

For everyday consumers and business owners, moving the pipeline directly to audio while keeping hardware footprints small enough for a laptop is a big deal. Block runs Square, which already handles order-taking for thousands of small restaurants. If you can deploy a responsive, fully conversational agent locally without running up massive cloud token bills, the unit economics of customer service change overnight.

The catch, however, is reliability. Because Dessa learns purely from conversational audio rather than vast scraped text repositories, its broad factual recall and complex reasoning could lag behind traditional LLMs. More critically, an AI voice that nails human breathing, hesitations, and crying fits creates a massive fraud surface. Keeping it locked behind closed doors for safety testing is a smart, necessary precaution—convincing people on the phone that a bot is human just got dangerously easy.