On October 23, 2026, Block's AI research lab unveiled Dessa, a full-duplex speech-to-speech model with a claim that immediately set it apart from every other voice AI announced this year: it was trained from scratch on raw human conversation recordings, with no text anywhere in its training pipeline.
No text-to-speech pre-training. No language model backbone. No transcripts. Just hours of recorded human conversation fed directly into the model. Block says Dessa learned to listen and respond the way people actually do — interrupting, backchanneling, overlapping — because that's the only kind of signal it ever saw.
What full-duplex actually means
Most voice AI systems are half-duplex: they wait for you to stop talking before they respond. Full-duplex models can listen and speak simultaneously, handling the natural overlap and turn-taking of real conversation. It's a harder problem, and it's why products like GPT-4o Voice or Google's Gemini Live still feel slightly robotic in their timing — they're grafted together from components that weren't designed for simultaneous input and output.
Dessa was apparently designed for full-duplex from the start, which is the architectural bet Block is making here.
The training story
According to Block's published details, training ran on a single NVIDIA GB200 NVL72 rack — a dense compute unit that packs 72 Blackwell GPUs into a liquid-cooled form factor. That's meaningful context: it's not a cheap experiment, but it's also not the multi-datacenter sprawl of a frontier model training run. The implication is that text-free speech modeling may be more compute-efficient than the current dominant approach of fine-tuning an LLM for voice.
Block also claims the resulting model runs locally, in real time, on a MacBook Pro M4 Max or better — inference faster than the pace of human speech. That's a striking claim for a model of this type, and independent researchers haven't yet had the chance to reproduce it.
What's missing: benchmarks
Block published more than a dozen conversation recordings as demonstration material. What it didn't publish was benchmark results. One analysis from Orca Router is blunt about this: everything we know right now is vendor claims, not measured performance. No word error rates, no latency numbers against comparable systems, no third-party evaluations.
That's not unusual for a research preview — it often takes weeks after an announcement for the research community to get access and start poking at a model. But it's worth keeping in mind before treating Dessa as a proven step forward rather than an interesting hypothesis.
Why Block?
Block (formerly Square) is a financial technology company, not an AI lab. Its interest in conversational speech AI makes some sense if you imagine Dessa eventually powering customer service interactions or the kind of voice-driven financial tools that would live inside Cash App or Square hardware. But Block hasn't said what it plans to do with Dessa commercially. For now it's a research preview with a public waitlist and a small group of AI safety researchers who have early access.
The text-free training approach is the thing worth watching. If it holds up under scrutiny — if the recordings are as good as the demos suggest and independent researchers confirm the local inference claims — it would represent a genuinely different path toward conversational AI than the one the industry has been walking since GPT-2.