Block’s Dessa explores conversation learned directly from human speech

The research model listens and speaks at once; the public material shows demonstrations and an early-access path for safety researchers.

Block’s research team has presented Dessa, a conversation model it says was trained entirely on human speech, without text or a pretrained model. The October 23 introduction explores an approach in which timing, interruptions and vocal expression remain part of the model’s input and output.

The official presentation describes Dessa as full-duplex: it can listen and speak simultaneously. Block says it runs locally on a MacBook Pro with an M4 Max or better. These are developer statements, not performance measurements from an independent hardware test.

Beyond a transcript

Speech contains pauses, emphasis and other signals that a transcript can lose. Dessa’s demonstrations explore how a model trained on sound can respond within that richer interaction, including when it begins speaking.

A natural-sounding response should not be taken as evidence that a system has feelings or reliably understands a person’s circumstances. Conversational presentation and factual dependability are separate questions.

A compact research effort

Block says it trained Dessa on a single NVIDIA GB200 NVL72 system. Its examples include overlapping conversation, changes in loudness, laughter and backchannel responses. Those clips demonstrate the kinds of behaviour the team is exploring; they do not establish performance across all speakers or conditions.

The official page invites a small group of safety researchers to request early access. It does not establish a generally available API or downloadable weights package. That access boundary makes Dessa a research introduction rather than a ready procurement choice. The next evidence to look for is how reliably the model handles misunderstandings and difficult conversations outside selected samples.

Sources

- [Dessa research presentation](https://block.xyz/dessa) — Block - [Block speech-model introduction date](https://techcrunch.com/2026/10/24/blocks-new-voice-model-sounds-eerily-human/) — TechCrunch — date corroboration only