Dessa: a voice model that learned to talk without ever reading
Most voice AI today is a relay race. Your voice gets turned into text, the text goes to a language model, the answer gets turned back into speech. Every handoff loses something: tone, timing, laughter, the sigh that tells you a customer is frustrated before they say a word. It is also why most AI phone voices feel a half-step behind the conversation.
Block, the company behind Square and Cash App, just showed a different way to build it. Their research lab announced Dessa, a voice model trained from scratch on nothing but recordings of human conversation. No text. No pretrained language model underneath. It learned to hold a conversation the way babies do, by listening.
What makes Dessa different
- Full-duplex. It listens and talks at the same time, like a person on a phone call. It can backchannel with a "yeah" or "got it" while you are still talking, and it handles interruptions without the awkward stop-start of current systems.
- Emotion is native. Because nothing gets flattened into text, Dessa hears and produces sarcasm, laughter, frustration, and relief. Block's demos show it laughing mid-sentence and finishing a speaker's sentence when they blank on a movie title.
- It runs locally. Block says inference runs faster than real time on a MacBook Pro with an M4 Max chip or better. No cloud round trip, which means lower latency and conversations that stay on the machine.
- It was trained on one rack. The whole model was trained on a single NVIDIA GB200 NVL72 system, learning language from hundreds of times fewer words than a typical LLM sees.
The team behind it is worth noting: a small lab inside Block, founded by former members of Dessa, the Toronto deep-learning startup Block acquired, led by Ragavan Thurairatnam and reporting directly to Jack Dorsey.
What it is not
Honesty matters more than hype here, so the caveats:
- You cannot use it yet. Dessa is a research preview. There is no API, no download, no pricing, and no announced release date.
- The demos are Block's own recordings. They are impressive, and Block says nobody steers the model during them, but there are no published benchmarks or independent tests yet.
- Access is safety-first. Block is opening early access to a small group of AI safety researchers before anyone else, specifically because a voice this convincing raises real impersonation and trust questions. We have applied for that early access program ourselves, because this is exactly the frontier our clients' phone systems will live on.
Why a local business should care
If your business runs on phone calls, and most service businesses do, the quality bar for an AI receptionist is not "can it take a message." It is "does the caller feel heard." A model that can hear a customer is in pain, in a hurry, or about to hang up, and respond like it, is a different product than a speech-to-text pipeline with a nice voice.
That day is not here yet. Dessa is a preview, not a product. But the direction is clear: voice AI is moving from reading scripts to actually listening. The businesses that already have their front office wired up, calls captured, follow-up automatic, will be the ones ready to plug a model like this in the week it ships.
That wiring is the work we do every day.