Block just proved you can train an AI entirely on human speech instead of text — and the result sounds like someone who learned English by eavesdropping at coffee shops.
The Summary
- Block released Dessa, an AI model trained exclusively on human conversation audio rather than text transcripts, marking a shift from text-first to speech-native AI development
- The model handles full-duplex conversation (both parties can talk simultaneously) but displays rough edges: mixing up names, awkward pauses, conversational hiccups
- Speech-first training could unlock AI that understands tone, interruption, hesitation — the actual texture of how humans communicate, not just what they say
The Signal
Every major AI model you've used — GPT, Claude, Gemini — learned language by reading the internet. Text first, speech second. Block's Dessa flips that. The model trained from scratch on audio of actual human conversations, no text transcripts involved. That's not just a technical curiosity. It's a different approach to teaching machines how humans actually talk.
The demos Block shared tell the story. In one conversation posted by Dessa lead Ragavan Thurairatnam, the AI maintains natural turn-taking and conversational rhythm. It also calls someone "Ronaldo" when that's not their name. It stumbles. It hesitates in ways that feel weirdly human but also uncanny.
"The model initially mistakes the speaker for somebody named 'Ronaldo' — the kind of error that reveals both the promise and limits of speech-first training."
Here's why the approach matters: when you train on text, you lose everything that isn't words. Sarcasm. Emphasis. The pause before someone admits they're wrong. The speed-up when they're excited. The filler words that signal uncertainty. Speech-first training captures that paralinguistic layer. Dessa isn't just learning what people say — it's learning how they say it, including the mess.
Block positioning this as a payments and commerce play makes sense. Voice is the interface for the agent economy. If your AI assistant needs to negotiate with a customer service bot, book a reservation, or handle a transaction over a call, it can't sound like a text-to-speech engine. It needs to handle interruptions, pick up on tone, and adapt in real time. Full-duplex conversation — where both sides can talk at once, like actual humans — is table stakes.
Key technical boundaries Dessa is testing:
- Can speech-only training reach the reasoning depth of text-trained models?
- Does training on conversational audio create models that generalize to formal speech, accents, background noise?
- Can you scale speech data collection the way OpenAI scaled text scraping?
The awkwardness in Dessa's demos isn't a bug to fix — it's signal about what's still hard. Name recognition, context retention across turns, knowing when to speak and when to shut up. These are problems text-based models skip because written language already solved them. Speech hasn't. Every transcription, every comma, every paragraph break is a human adding structure that doesn't exist in raw audio.
The Implication
If Dessa or something like it works, the agent economy gets cheaper and weirder. Cheaper because speech-native models could run on less compute than text-to-speech pipelines stacked on top of LLMs. Weirder because agents that learned from human conversation will inherit human conversational tics — the ums, the false starts, the moments where they sound uncertain even when they're right.
Watch what Block does next with this. If they integrate Dessa into Cash App or merchant tools, that's a signal they think speech-first AI is ready for production. If this stays a research demo, it means the technology isn't there yet. Either way, training AI on how people actually talk instead of how they write is the right direction. We don't text our thoughts. We speak them.