Block Dessa Speech AI: What Becomes Your Notes in 2026
A natural conversation still needs a careful written handoff.
October 24, 2026
Block's Dessa is a speech-to-speech AI model built for live conversation. It isn't a demonstrated replacement for a writing or note-taking app. If you want useful notes from a conversation, keep tentative ideas separate from decisions and review the written result. Block's launch page describes audio coming in and audio going out, with subtitles added afterward to its recorded demonstrations.[1]
That last detail deserves attention. A voice can carry hesitation that disappears in a tidy sentence. Put those sentences in a project brief and a half-formed idea can start looking like an instruction.
Dessa drew fresh attention on October 24. At this morning's check, its Hacker News story was on the front page with 423 points and 160 comments. Those are discussion counts, not user numbers or proof of quality.[4] Jay Peters covered the October 23 announcement for The Verge. I read that report through an October 23 archive capture because the live page rejected the fetch.[8]
What Block's Dessa announcement establishes
According to Block, Dessa learned from human speech alone, without text or pretrained models. It can listen and speak simultaneously, the behavior called full-duplex. The company says it can run locally on a MacBook Pro M4 Max or better. These are vendor claims, not results of a DictaFlow benchmark.[1]
The launch page offers recordings and a waitlist, with early access for a small group of safety researchers. That is a reason to watch the work, not a reason to replace your daily documentation setup today.[1]
Peters found the recorded exchanges mostly natural but sometimes awkward. He noted one conversation where Dessa initially mistook the person for someone named Ronaldo. His report also described an unexpected film-related tangent.[8] That is a useful check on the launch's enthusiasm. Fluent conversation can include a wrong assumption.
I haven't tested Dessa. The recordings and reporting tell us something about the public preview. They don't establish how it would handle your colleague's name, a noisy desk, or a long planning session.
A good conversation can leave a bad project note
Imagine a podcast producer planning an interview. This is a fictional example, not a Dessa test.
The producer says, "Maybe open with the customer story. I'm not sure we have permission yet." A collaborator answers, "Yeah, that could work." They spend another minute talking about music and move on.
A later note reading "Open with the customer story" would be wrong. The condition about permission belongs in the note. So does the uncertainty. The collaborator expressed interest, without resolving either issue.
Even a perfect transcript leaves someone to decide what the exchange means for the production schedule. Should the editor start cutting? Who checks permission? Is there a backup opening? A transcript can preserve every word and still leave those questions unanswered.
This is why I want two different outputs from a voice tool. During discussion, it should let people explore unfinished ideas. When preparing a handoff, it should preserve the unfinished parts visibly enough that another person won't mistake them for instructions.
Block is right to draw attention to information carried by sound. Its page points to stress, pauses, and emotion as parts of speech that text can lose.[1] For writers, the useful follow-up is deciding which of those signals need an explicit sentence when the recording becomes a document.
"Permission pending" would do real work in this example. A cheerful voice wouldn't settle it.
Try a handoff exercise before trusting the notes
If you get access to Dessa, or evaluate another conversational tool, use a short made-up planning discussion. Avoid private client material. Keep the exercise small enough that you can compare the result with what was said.
Choose one job, such as drafting instructions for a podcast editor. Speak a firm decision about the episode length. Then suggest an opening you haven't approved. Include an unfamiliar guest name and leave one question unresolved.
At the end, ask for a written handoff if the product supports that output. If it doesn't, write your own note from the discussion. That distinction matters: this is a workflow exercise, not a claim that Dessa currently exports notes or connects to your editor.
Look for specific meaning changes:
- Did the proposed opening become an approved edit?
- Did the note keep the unresolved permission question?
- Can the next person tell which episode-length instruction you settled on?
- Is the guest's name accurate, or does it need checking?
Don't award a pass because the note sounds professional. Read it as the editor who receives it tomorrow, without the voice recording or your memory of the call. Underline anything that could send that person into work you didn't authorize.
A short note can pass. A long one can fail. For this exercise, preserving the status of the proposed opening matters more than preserving the exact wording of every casual remark.
I would also check whether you can correct the written handoff directly. If the only way to fix one name is another extended conversation, that adds friction to a task that should take seconds. This is a buying question to investigate, not a defect established by the Dessa preview.
Sometimes the next step is simply dictating the brief
You might enjoy an open-ended voice discussion and still prefer to dictate the final handoff yourself. After talking, open the editor's brief and say the instructions you intend to stand behind. Put unresolved items in their own paragraph.
For the fictional podcast, that could be: "Keep the episode under thirty minutes. The customer-story opening is a proposal. Don't cut it in until we confirm permission. I still need to check the guest's surname."
Then read it. Check that the recipient can act on the firm instruction while leaving the proposal alone.
DictaFlow supports that text-first step: hold a shortcut, speak, and place text into a supported field. Its current product page describes dictation for emails, documents, and AI tools, with review before sending.[6] I build DictaFlow, so that recommendation has a commercial connection. It is a suggestion for preparing editable text, not a claim of a Dessa integration or automatic verification.
Dessa's speech-first research is worth following. My practical question for a writing workflow comes after the interesting conversation: can the person receiving the brief see that the opening still needs permission? Leave that sentence in the document until someone resolves it.
Sources
[1] Block: Dessa project and access details
[4] Hacker News front page, captured October 24 at 10:22 EDT