Block unveils Dessa, a voice model it says was trained only on speech

On Friday, October 23, 2026, Block's small in-house AI research lab unveiled Dessa, a voice model that Block says was trained only on human speech. Block cofounder Jack Dorsey posted the announcement at 2:23 p.m. Eastern, calling Dessa “the most authentic full-duplex speech model ever built” (Block publishes no metric for “authentic”) and saying talking with it “feels like talking with a person… full of emotion, nuance, pauses, and a natural cadence.” Block describes the lab as founded by former members of Dessa, the Toronto deep-learning startup, and led by Ragavan Thurairatnam; it reports directly to Dorsey.

Full-duplex means Dessa listens while it talks, the way people do. It takes audio in and gives audio out — no transcription step in between — and Block says it can cut in, backchannel with a “yeah!” or “got it!,” and hear emotion, sarcasm, and laughter that a text pipeline loses. Block says inference runs faster than real time; it has published no latency figures.

Block's pitch is that routing speech through text loses timing, tone, and emotion, and that many voice assistants wait for you to finish before they answer. Block compares Dessa's training to how babies learn to talk without ever seeing a written word, and says no text was used in training.

Block argues this is a harder learning problem than it sounds. Speech never repeats: say a sentence twice exactly the same way and the data still comes out different, while typed text is identical no matter who types it. A single spoken word takes thousands of times more bytes than the same word in text, and stressing a different word changes the meaning — Block's example is saying “I didn't say she stole my CUDA code” seven ways. Block says Dessa learned language from hundreds of times fewer words than large language models need, a figure it gives without supporting data, and that the team trained it on a single NVIDIA GB200 NVL72 rack. NVIDIA appears only as the hardware named; Block's page describes no partnership or endorsement.

Block published recorded conversations it describes as live, with no one steering the model. Block's descriptions of the clips say Dessa laughs mid-sentence when something is funny, gets louder when asked to speak up, finishes a person's sentence when they blank on a movie title, and gasps under its breath at a surprise announcement before answering with what Block says sounds like teary gratitude. One recording has Dessa talking with OpenAI's GPT-Live-1; Block notes that model's developer is not affiliated with Block and has not endorsed Dessa.

Block has not published a parameter count, API, pricing, weights, or standard benchmarks. The Verge, which covered the launch, found the demos mostly natural but occasionally awkward — in one, Dessa mistakes the speaker for someone named “Ronaldo.” Block's “world's first” claim also carries its own qualifier: “as far as we know.” TechCrunch notes that OpenAI's GPT-Live, released earlier this year, is also a full-duplex model; Block's distinction is that Dessa was trained on speech alone, with no text or pretrained model. That claim has not been independently verified. CoinDesk reported mixed reactions on X, from praise for how human the demos sound to warnings that such voices could make phone scams harder to detect.

Block gives three reasons: curiosity, better understanding between humans and AI, and safety. Its safety argument is that AI keeps gaining power and independence while staying nearly deaf to the emotion in a speaker's voice — imagine being in serious pain while a voice AI has no idea and just keeps going. Dessa is a step toward fixing that, in Block's telling.

For now, there is no public access. Block's page says launch links are coming soon, and Block says it is giving a small group of AI safety researchers early access to find risks the team hasn't anticipated. Some secondary reports describe a public waitlist, but the page's own “Join the waitlist” button is hidden and not yet active. What to watch: whether broader access follows, whether independent researchers get to test the model, and whether a speech-only model can learn facts and new languages as efficiently as text models can. Block lists those as open research questions, not solved problems.