Block Unveils Dessa, a Full-Duplex Speech Model Trained Only on Speech

Most voice assistants today are a relay race. Speech gets turned into text, a language model answers in text, and a synthesizer reads it back. Block's AI lab says its new research preview, Dessa, skips all of that. It is a speech-to-speech model that listens and talks at the same time, and Block says it was trained only on human conversation audio, with no text and no pretrained model as a starting point.

Two claims stand out. Block says training took a single NVIDIA GB200 NVL72 rack, 72 GPUs, compared with the 1,016 H100s Kyutai lists for its open Moshi model. And Block says Dessa runs locally, faster than real time, on a MacBook Pro M4 Max or better. If those hold up, natural voice AI gets cheaper to build and less dependent on the cloud.

We like the direction here. The text-first voice stack loses a lot of what makes speech human: the laugh, the sigh, the sarcastic tone, the well-timed 'yeah' that tells someone you're still with them. Block's argument is that a model raised on speech keeps those signals instead of flattening them into words. The demo clips show backchanneling, mid-sentence laughter, and natural interruptions, and Block says none of them were steered.

The local angle matters just as much. A voice model that runs on a laptop means conversations that never have to leave the device. For privacy, for latency, and for anyone building voice tools without a big cloud bill, that is a genuinely useful shift.

Now the honest part. This is a demo, not a result. Block has published no benchmark scores, no parameter count, no architecture, no paper, and no weights. Access is a waitlist. Compare that with Moshi and NVIDIA's PersonaPlex, which ship open weights and published latency numbers. Until Dessa is measured on shared tests, the efficiency claims are Block's word, and curated demo clips are the easiest thing in AI to get right.

There is also a downside worth naming if this scales. A voice that laughs, interrupts, and responds to tone the way a person does is more persuasive and harder to tell apart from a human. A model that learned its manners from real conversations will be very good at sounding trustworthy. That is great for accessibility and customer support, and a real risk in scam calls and impersonation. I'm watching whether Block pairs any release with disclosure or watermarking plans.

Our read: the speech-only approach is a serious idea and a welcome challenge to the text-first default. Show us the numbers.

Full-duplex means a model can hear and speak simultaneously, rather than waiting for a turn to end. Kyutai's Moshi, the best-known open full-duplex model, starts from a text language model backbone and predicts text tokens alongside audio. OpenAI's GPT-Live-1 is a closed, cloud-based voice layer; one of Block's demo clips shows Dessa in conversation with it.

Block's lab notes why speech-only training is hard: spoken words never repeat exactly, and a word in audio takes thousands of times more data than the same word in text.