Block Unveils Dessa, a Full-Duplex Speech Model Trained Purely on Human Conversation
Block's AI research lab has released Dessa, a full-duplex speech model trained exclusively on human conversation. The company announced the model on October 23, 2026, as a research preview. Dessa is a speech-to-speech model designed to listen and talk at the same time. Notably, Block says it used no text data and no pretrained models—a marked departure from the text-first stacks that underpin most voice agents on the market today.
TL;DR
- Runs on: A MacBook Pro M4 Max or better, locally, faster than real time.
- Performance: Demo-only so far. Live recordings demonstrate backchannels, mid-sentence laughter, and natural interruptions.
- Best: Trained on a single NVIDIA GB200 NVL72 rack (72 GPUs). Moshi, for comparison, used 1,016 H100 GPUs.
- Worst: Zero published benchmark scores. No weights, paper, API, or pricing has been released.
What Is Block's Dessa?
Dessa is a true speech-to-speech model: audio in, audio out. According to Block's project page, everything it knows came from human speech.
Because it is full-duplex, Dessa listens while it speaks—enabling the kinds of overlapping, responsive exchanges that define natural human conversation. This stands in contrast to the turn-based "listen, then respond" architecture typical of conventional voice assistants.
Training Efficiency and the 2026 Context
Dessa's training footprint is a notable talking point. Block reports that the model was trained on a single NVIDIA GB200 NVL72 rack—72 GPUs in total—compared with the 1,016 H100 GPUs used to train Moshi, a comparable full-duplex speech model. In an era where frontier model training costs are under intense scrutiny, this gap in compute efficiency is significant, though it remains unclear how directly the two projects' training regimes can be compared.
What Block Has and Hasn't Shown
So far, Dessa is a demo-only research preview. Public live recordings highlight the model's ability to produce backchannels (such as "mm-hmm" acknowledgments), mid-sentence laughter, and natural interruptions—behaviors that are notoriously difficult to replicate in pipeline-based text-first systems.
What Block has not released is equally important: there are currently no published benchmark scores, no model weights, no technical paper, and no API or pricing information. As a result, Dessa's capabilities cannot yet be independently verified or compared against established speech models.
Why It Matters
Dessa's speech-only, full-duplex approach represents a distinct design philosophy in voice AI. Rather than converting audio to text, reasoning over it, and converting back to audio, Dessa works entirely in the speech domain. If the approach proves robust beyond demos, it could point toward more natural, lower-latency conversational agents—and a substantially more compute-efficient path to training them.
via MarkTechPost