News · 2026-10-24

Block previews Dessa, a conversation model it says learned without text

Block has previewed Dessa, a speech-to-speech conversation model it says was trained from scratch on human conversation audio, without text data or a pretrained model. The public release offers selected recordings and a safety-researcher access form, rather than a downloadable model or a general public service. Its central research question is whether learning directly from sound produces better conversational timing and emotional expression.

Key facts

• Block describes training on one NVIDIA GB200 NVL72 system, a rack containing 72 graphics processors.
• The release was live by October 24; its Hacker News discussion appeared October 23, while the release page is undated.
• Block says Dessa can listen while speaking and run faster than real time on an M4 Max MacBook Pro or better.
• Primary source: Block’s Dessa release page.

The impressive part of the pitch is easy to grasp: speech contains more than words. A pause can signal uncertainty, a rising pitch can turn a statement into a question, and a short acknowledgment can show that a listener is following along. A system that turns sound into a transcript, reasons over text, and then reads out an answer can lose some of that information between stages. Block says Dessa instead learns from “nothing but human conversation.”

Think of the difference between reading the script of a play and listening to the actors rehearse. The words may match, but the timing, emphasis, hesitation, and interruptions carry additional meaning. Direct acoustic learning could capture those patterns together. That is a useful explanation of the research motivation, not a disclosed account of Dessa’s implementation: Block has not specified the audio representation, network design, training objective, corpus size, or division of work among components.

The distinction from existing assistants needs care. Full duplex—listening and speaking at the same time—is not unique to Dessa. OpenAI’s GPT-Live announcement describes continuous hearing and speaking, with harder reasoning delegated to a frontier model. Block’s narrower novelty claim is the absence of text training and a pretrained foundation. Our full-duplex speech lesson explains why simultaneous input and output is a separate design choice from training provenance.

Hardware language also changes the scale of the story. Block says training used one GB200 NVL72, while NVIDIA’s specifications identify that product as a rack-scale system with 72 Blackwell graphics processors and 36 Grace processors. One system is a meaningful engineering constraint, but it is far removed from training on one ordinary graphics card. Block’s claim of learning from hundreds of times fewer words than text models lacks a published counting method and matched comparison, so it cannot yet be audited.

The recordings show or describe laughter, short acknowledgments, expressive intonation, responses to spoken requests, and a movie-title completion. Block says they are live conversations, with subtitles added afterward. Selected clips establish a demonstration, not a general accuracy rate or a fair comparison with another assistant. A plausible emotional response does not prove reliable emotion recognition, and recalling one title does not establish broad factual competence.

Block’s safety motivation is that distress and other vocal cues can disappear when sound is flattened into text. That is a reasonable question to investigate. It is also a source of risk: humanlike reactions can make a system feel more understanding than its tested capabilities warrant. The release includes no controlled study of emotion-recognition accuracy, user consent, privacy, or the effects of apparent empathy. The difference between an expressive voice and dependable judgment matters particularly when people are vulnerable.

Early listeners disagreed about the same behaviors. In the Hacker News discussion, some praised cadence and local-inference potential; others heard overacting, awkward vocal affect, or too many interruptions. The checked snapshot had 384 points and 138 comments. Reddit’s launch discussion likewise mixed enthusiasm with questions about intelligence and availability. Those are self-selected reactions, not an expert assessment.

The strongest counterargument is that acoustic naturalness may be a narrow strength best paired with a separate reasoning model. Block’s own clips include a conversation with GPT-Live-1, but the release does not demonstrate a general tool-calling system, text-control interface, or coding capability. A hybrid could be useful; its intelligence would come from the combined system rather than Dessa alone.

For now, there is no published weight download size, and Block states no numerical minimum GPU-memory requirement. Its M4 Max claim is a hardware-class statement without a specified memory configuration; it does not establish that every machine bearing that chip can run the model. Dessa is a substantive research preview, with a testable direction and unusually vivid demonstrations. Public access, implementation disclosure, and controlled evaluation are the evidence still needed to turn the preview into a dependable tool.

Primary source, verified: read the paper →

Ground Truth. (2026, October 24). Block previews Dessa, a conversation model it says learned without text. Ground Truth. https://groundtruth.day/news/block-dessa-speech-only-conversation-preview.html

@misc{groundtruth:block-dessa-speech-only-conversation-preview,
  title  = {Block previews Dessa, a conversation model it says learned without text},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {oct},
  url    = {https://groundtruth.day/news/block-dessa-speech-only-conversation-preview.html}
}

Topics: voice-ai · speech · block · research-preview · human-ai-interaction