Block releases Dessa, a full-duplex speech model it says was trained only on human conversation

Block's AI research lab has published a research preview of Dessa, which the company describes as a full-duplex speech model trained only on speech. On its announcement page Block says the model was trained from scratch on human conversation, with no text and no pretrained models, and that it takes audio in and produces audio out directly. Block says Dessa listens and talks at the same time and learned when to interrupt and when to backchannel, and that it learned language from "hundreds of times fewer words" than large language models use, without giving an exact figure.

On performance, Block states that inference runs faster than real time and that the model runs locally on a MacBook Pro M4 Max or better. It says training ran on a single NVIDIA GB200 NVL72. Availability is limited: the page describes a research preview, says launch links are coming soon, and states that a small group of AI safety researchers has been given early access to look for unanticipated risks. Block frames emotion perception as a safety question, arguing that a voice AI that cannot hear emotion is itself a safety problem.

Why it matters for digital humans: Full-duplex speech is the layer most conversational avatar products are still missing. Today a typical digital human pipeline runs speech recognition, then a text LLM, then text-to-speech, then lip-sync — a chain that discards prosody, sarcasm and laughter at the first step and introduces turn-taking delay at the last. A speech-native model that handles overlap and backchannels natively would change what an avatar can do while listening, which is where most interactive avatars still look dead. Two caveats: every performance claim here is Block's own, with no independent benchmarks or absolute latency figures published, and Dessa is audio-only — pairing it with a face model is a separate and unsolved latency budget. There is also no public availability yet, so this is a direction of travel rather than something to build on.