Block Unveils Dessa: AI Model Trained Solely on Human Conversation
Block, the company behind Cash App and Square, announced Dessa this week — an artificial intelligence model for speech processing that the company calls "the most authentic full-duplex speech model ever built." Unlike almost every current AI voice system, Dessa was trained from scratch exclusively on recordings of human conversation: no internet text, no pretrained models, no text-intermediate pipeline.
The approach: speech instead of text
The conventional pipeline for voice assistants follows a three-step path: the user speaks, the system converts audio to text through speech recognition (ASR), a language model processes the text and generates a response, and finally that response is synthesized into audio through a text-to-speech system. Block considers this flow artificial and severely limiting.
"We believe voice is the most natural way to interact with intelligence," the project team wrote on Dessa's page. "Treating everything as text and then throwing away information that text cannot carry is ridiculous, worsens the experience, and is potentially dangerous."
Dessa operates differently. It is a genuine speech-to-speech model: audio goes in, audio comes out. Trained "from scratch on nothing but human conversation," the model learned to hold a real-time conversation simply by listening. There is no intermediate translation to text. Inference runs faster than real time on a MacBook Pro M4 Max or higher.
What sets Dessa apart technically is its native full-duplex capability. The model listens and speaks simultaneously, produces natural backchannel sounds ("yeah," "got it," "uh-huh") at the right moment, and can interrupt conversation naturally — exactly as a human conversation works. Additionally, it natively processes emotion, sarcasm, laughter, and sighs without needing additional emotion-analysis layers.
The team behind it: the forgotten lab
Behind Dessa sits a small AI research lab inside Block, founded by former members of Dessa — the Dessa, the Toronto deep-learning startup that originally produced realistic synthesized voices for advertising. The team has been building deep learning models since 2012 and generative speech models for over eight years. According to its own description on the project site, the team "chases frontier problems others ignore" and reports directly to Jack Dorsey, co-founder and chairman of Block.
The connection to Buzz — the team chat platform launched in July 2026 that allows humans and AI agents to share channels — is evident. Dessa complements Dorsey's strategy of building natural interfaces between humans and machines, where conversation replaces the keyboard as the primary means of interaction.
Business context: Block's AI bet
Dessa's announcement fits into a moment of radical transformation at Block. In February 2026, the company laid off approximately 4,000 employees — more than 40% of its workforce — arguing that artificial intelligence would enable it to operate faster and more efficiently. Since then, the company has reported tripling the number of products developed internally.
In the second quarter of 2026, Block's gross profit rose 27% year-over-year to US$2.9 billion, with revenue of US$6.1 billion. Dorsey attributed significant portions of that result to AI, saying in a shareholder letter: "Internally, AI is helping us move faster and improve quality. Externally, it's helping us build products that act earlier for customers and sellers."
The company already has two AI assistants in Cash App — Moneybot and Managerbot — and is exploring pricing models for AI tools before public launch. Dessa represents the next layer of that strategy: rather than building on third-party APIs or existing language models, Block is investing in proprietary models trained from scratch, aiming to differentiate its AI stack from competitors that rely on OpenAI, Anthropic, or Google.
Why this matters
The premise of Dessa — that human conversation contains all the information necessary to build conversational intelligence — is ambitious and controversial. Most modern AI labs rely on internet-scale text corpora as the foundation of knowledge, applying alignment techniques afterward to make models "friendly." Block's approach flips this: what matters in conversation is not factual knowledge but rhythm, emotion, and social interaction.
There are risks to this approach. A model trained only on human conversations will naturally have whatever factual knowledge people express in conversation — which means gaps in technical, scientific, and specialized areas that rarely appear in everyday dialogue. The project itself admits it is giving early access to a small group of safety researchers to "find the risks we haven't anticipated."
What Dessa reveals, perhaps more importantly, is that Block believes differentiation in the next wave of AI will not come from bigger models or larger datasets, but from how intelligence is expressed. Voice, with its emotional and temporal load, is where the next leap happens — and the company is willing to build the model from scratch to claim it.
Sources: The Verge, AICoder, Unite.AI, Payments Dive, AGTP Insights
✓ Independent sources cross-checked and verified before publishing
