The Pulse
Block Trained Dessa on Human Speech and One GB200 NVL72
Block says Dessa learned to hold real-time conversations from human speech alone, using one NVIDIA GB200 NVL72. The speech-to-speech model is being offered to a small group of AI safety researchers, while Block’s examples include both expre
AI.info Team · Written with AI · fact-checked ·
Block says it trained Dessa, its new conversational AI, on a single NVIDIA GB200 NVL72 using human speech alone. The company says the model learned without text or pretrained models, then built on that audio training to speak and listen at the same time. Block announced the system on October 23, 2026, and is offering early access to a small group of AI safety researchers.
Dessa skips the text transcript
Many voice assistants route speech through a written transcript before generating a spoken response. Block says Dessa instead takes audio in and produces audio out, preserving cues such as emotion, sarcasm, laughter and pauses that can be lost when speech becomes text. Its full-duplex design lets it listen while speaking, so it can respond to interruptions and conversational signals without waiting for a speaker to finish.
“We trained Dessa from scratch on nothing but human conversation. Sounds! No text, no pretrained models,” the Dessa team writes on its announcement page. Block says inference runs faster than real time and that the model runs locally on a MacBook Pro M4 Max or better. The company has not made general access available; its site has a waitlist and a researcher access request, while its initial access offer is aimed at researchers studying safety.
The demos show fluency—and mistakes
Block’s examples include a conversation about films, a guessing game and a surprise-trip story. In one recording, Dessa reacts with laughter and emotional inflection; in another, the model responds to someone who cannot recall a movie title. The company presents the clips as live conversations in which the person does not steer the model’s replies.
The early exchanges are not uniformly smooth. The Verge describes one conversation shared by Dessa lead Ragavan Thurairatnam in which the model initially mistakes the speaker for “Ronaldo,” while noting that the back-and-forth is mostly natural. Separately, The Verge points to Block’s Christopher Nolan films example, where Dessa goes on an unexpected tangent about Bryan Singer’s Superman films. Such examples show why convincing timing and expressive speech do not, by themselves, establish that a model understands a conversation reliably.
Block frames speech as a safety issue
The company says it built Dessa for curiosity, better understanding between people and AI, and safety. Its argument is that a system that detects distress in a person’s voice could respond differently from one that only processes the words in a transcript. Block offers the example of someone in serious pain whose voice assistant fails to recognize the emotion and keeps going.
“We’re giving a small group of AI safety researchers early access to find the risks we haven’t anticipated.”
— Ragavan Thurairatnam, Dessa lead, writes on Block’s page.
That is a stated motivation, not evidence that Dessa can reliably identify distress or respond appropriately. Block says it wants outside researchers to look for risks the team has not anticipated before broader access. The company also describes the model as new territory for research, including questions about how efficiently it learns language and facts from sound, and whether it can learn additional languages.
Language learned from sound alone
Block’s technical claim is narrower than saying speech-only models are new: it says that, as far as the team knows, Dessa is the first AI that can hold a realistic live conversation after training only on human speech. The company compares the amount of language data it learned from with the much larger word counts used by language models, while acknowledging that audio carries far more information than text.
The claims now face practical tests beyond the selected demos: whether Dessa retains facts, handles unfamiliar accents and languages, and recognizes emotional cues without misreading them. Block’s early access program puts those questions in front of safety researchers first; the model’s broader availability is not yet open, with the company’s site offering a waitlist.