Block introduces Dessa speech model

Block’s in-house AI research lab announced Dessa on October 23, 2026.[1] Block describes it as a full-duplex speech-to-speech model, meaning it listens and speaks at the same time.[2] The company says it trained the model from scratch on human conversation alone, without text or pretrained models.[2] Block says that, as far as it knows, Dessa is the first AI to hold a realistic live conversation after training only on human speech.[2] It calls Dessa “the most authentic full-duplex speech model ever built.”[2] These are Block’s characterizations of its model.[2]

Block says Dessa natively hears and speaks emotion, sarcasm, laughter, frustrated sighs and sadness in a voice.[2] It also says the model learned from real conversation when to interrupt and when to backchannel.[2] Block reports that inference runs faster than real time and that the model runs locally on a MacBook Pro M4 Max or better.[2] The company says it trained Dessa on one NVIDIA GB200 NVL72 and that the model learned language from hundreds of times fewer words.[2] Those details give teams concrete claims to test against their own hardware and conversational requirements.

Block says its posted conversation clips are live recordings, nobody steers the model during them, and Dessa chooses what to say, when to say it and how to say it.[2] Subtitles were added after the recordings.[2] In Block’s examples, Dessa laughs mid-sentence when something is funny, gets louder when asked to speak up and finishes a speaker’s sentence when the speaker cannot recall a movie title.[2] The Verge described the shared conversations as occasionally awkward and uncanny.[3] The clips offer examples of the behavior Block highlights, but the evidence here does not establish how the model performs beyond those demonstrations.

Block says it built Dessa out of curiosity, to improve understanding between humans and AI, and for safety.[2] The company argues that voice AI can miss signs of distress because it is nearly deaf to emotion in people’s voices.[2] It is offering early access to a small group of AI safety researchers to look for risks the team has not anticipated, and its page invites visitors to join a waitlist.[2] AICoder reported that Dessa had no public API, weights or download at the time of its article.[4] That means teams interested in hands-on testing should distinguish the waitlist and limited early access from a generally available integration.