The moment a voice responds can matter as much as the words it chooses. A brief acknowledgement while someone is still talking carries a different rhythm from waiting until they finish. Dessa, a research preview from a small team at Block, puts that timing at the centre of its pitch.
On its official research page, Block describes an audio-in, audio-out model trained from scratch on human conversation. The announcement raises a useful question: what has the company reported, and what can readers actually access?
Three parts of the announcement
First comes training. Block says it used human conversation alone, with no text and no pretrained models. That is the company’s account of how Dessa learned; this article does not independently audit its training data. The statement should not be expanded into undocumented architecture details or a verdict that it outperforms every other voice system.
Second is interaction. Block describes Dessa as full-duplex, meaning it can listen and speak at the same time. The team points to interruptions and short acknowledgements as parts of the conversation, rather than treating every exchange as a neatly separated question and answer.
Third is execution. Block says inference runs faster than real time on a MacBook Pro M4 Max or better. It also says training used a single NVIDIA GB200 NVL72. Both are attributed hardware and performance claims, not results of a test conducted for this article.
Listen to the evidence category, too
The official page offers selected company recordings. They illustrate the behaviour Block wants to show, but they are not an independent benchmark. Expressive speech and a well-timed response do not, by themselves, establish reliable understanding of a person’s emotions.
We have not carried out a listening trial or a comparison against competing models. That matters when reading superlatives about realism: the company’s presentation cannot substitute for a shared evaluation method.
Local does not mean downloadable
Running a model on a laptop and distributing that model are separate decisions. The inspected page invites a small group of AI safety researchers to request early access and displays a coming-soon notice for launch links. It does not establish a public API, a downloadable release or a firm general-release date.
Readers should therefore understand Dessa as a research preview. The local-running claim describes what Block says its system can do, rather than a setup guide available to everyone today.
Jasper’s assessment
The useful thread to follow is conversational timing. A broader evaluation could examine interruptions, different speakers and the way mistakes are recorded. Those are proposed research questions, not experiments we have completed. For now, Dessa offers a concrete speech-first research direction, with the company’s demonstrations and access limits kept clearly in view.
