Explainer

Dessa AI TTS? Why Dessa Is Not Text-to-Speech at All

2026-11-05 · 7 min read · By DessaAI Daily

If you searched for “Dessa TTS”, you are probably looking for a text-to-speech tool — type something, get a voice reading it back. That is exactly what Dessa is not. Block’s Dessa, announced on October 23, 2026, is a speech-to-speech model with no text step in the middle at all. Understanding that difference is not pedantry: it is the whole reason Dessa exists, and it changes what you should — and should not — expect from it.

What TTS actually means

Text-to-speech, usually shortened to TTS, is the last stage of the traditional voice pipeline. First, speech recognition turns what you said into text. Then a language model, trained mostly on written words, works out a text reply. Finally, TTS reads that reply aloud in a synthetic voice. Each stage is separate, and each handoff loses something. Your tone, hesitation and sarcasm rarely survive transcription, and the reply voice has to fake feeling back in — often from written cues or tags that tell it where to laugh, whisper or pause.

Modern TTS can sound impressively human on a single sentence. The problem Block points to is conversation. A TTS voice is, at its core, reading. It waits for a finished text, then performs it. It was not listening while you spoke, and it cannot react to the crack in your voice halfway through a sentence, because in its world that crack was never data — it was stripped out before the thinking started. We break that pipeline down step by step in how Dessa AI works.

What Dessa does instead

Dessa takes audio in and produces audio out, directly. Block says it was trained from scratch on human conversation audio only — no text, no pretrained models. It did not learn to read and then learn to speak. It learned to talk the way Block argues babies do: by listening. Because its training data was real conversation, it learned far more than words. It learned accents, pauses, breathing, timing, sarcasm and emotion, because in speech all of that is part of the message.

That is why Block’s demos sound different even when the words are ordinary. Dessa laughs the moment something is funny, mid-sentence, rather than after a polite pause. It backchannels with “yeah!”, “got it!” and “ooh” while the other person is still talking. It knows when to cut in and when to let you finish, and Block says it gets louder when you ask it to speak up — a behaviour nobody explicitly taught it. None of that is TTS behaviour. It is full-duplex conversation behaviour, possible only because there is no text relay to wait for.

So is there a Dessa TTS you can use?

No. There is no Dessa text-to-speech product, no Dessa voice generator you can type into, no API that turns your script into Dessa audio, and no download. Block has announced no pricing, no API, no model weights and no public release date for Dessa in any form. It is a research preview. A waitlist is open on Block’s official page, and early access goes first to a small group of AI safety researchers. Anyone offering a “Dessa TTS” tool, voice pack or generator today is using the name, not Block’s model — a risk we cover in is Dessa AI safe?

If you need TTS today, that is a different shopping list. Established voice tools built around text — including the tag-based approach we compared in Dessa AI vs ElevenLabs — are available now, with published pricing and APIs. Dessa is not competing with them as a product yet, because Dessa is not a product yet. It is an experiment asking whether removing text entirely produces a more human conversation.

Why people keep searching for it anyway

The confusion is understandable. For twenty years, “AI voice” has meant TTS to most people: a computer voice is a voice reading text. Search habits follow that history, so “Dessa TTS” becomes the natural query even for people who mean “Dessa’s voice”. Block’s claim, in plain terms, is summarised in what Dessa AI is: the voice is not a layer on top of a text brain. The voice is the brain’s native language.

That distinction also explains Dessa’s known limits. A model that has never read a book, a manual or a webpage cannot be assumed to match a text-trained assistant on factual, document or tool-using work, and Block has published no benchmarks to claim otherwise. A TTS system inherits its text model’s knowledge. Dessa, by design, starts without it — which is precisely the research question its small Block lab, led by Ragavan Thurairatnam and reporting to Jack Dorsey, is testing.

The honest bottom line

If you came here for a Dessa TTS download, API key or voice to narrate your text: it does not exist, and Block has not announced one. If you came here because you heard Dessa’s demos and wondered what kind of voice technology that is: the answer is speech-to-speech, full-duplex, trained only on human conversation, running — in Block’s local-inference claim — faster than real time on a MacBook Pro with an M4 Max chip or better. Calling it TTS undersells what is novel about it and oversells what you can do with it today. Both errors are worth avoiding.

Sources: Block’s official Dessa page (block.xyz/dessa), re-checked November 5, 2026; The Verge and TechCrunch coverage, October 2026.

Frequently asked questions

← More Dessa AI articles Join the official waitlist ↗