Block's Dessa: now indistinguishable from a human

Full-duplex speech-to-speech: silence stops being a timeout

I pressed play and stopped looking for the seam

Dessa is a voice model Block announced today. Audio in, audio out, trained from scratch on human conversation only, without a single line of text. And in the published demos, as far as I can tell, it is indistinguishable from a person.

My method was not rigorous. I opened the Dessa page, read a bit, and played the recordings one after another. With voice demos I usually play a game: find the spot where the machine shows. The pause before the answer that is too clean. The pitch that rises the same way at the end of every sentence. The laugh that lands half a second after the joke. Here I gave up on the game early.

It laughs mid-sentence when something is funny. While the other person talks it drops in its "mm-hm", "yeah", "right", without taking the turn. It overlaps, stops, picks up again. It behaves like a person on a phone call, not like an assistant waiting for its slot.

The only slip I have seen reported comes at the start of one recording, where it calls the other speaker "Ronaldo" for no reason. I mention it because it is the most human mistake it could have made.

What we know about Dessa, in five lines

Block's page is linked above, and rewriting it would be pointless. The facts that matter to a developer:

• no text and no pretrained models: everything it knows comes from spoken conversation;
• full-duplex: it listens while it speaks, so it reacts to interruptions, pauses and laughter;
• faster than real time at inference, according to Block;
• runs locally on a MacBook Pro M4 Max or better;
• research preview: no API, no weights, early access for a small group of AI safety researchers, plus a waitlist.

Ragavan Thurairatnam, who leads Block's AI research lab, called it "a GPT3 moment for speech". That is launch-day copy. After the recordings, I get where it comes from.

Why today's voice agents sound like voice agents

If a client asks you for a voice assistant today, you almost certainly build a cascade. Speech-to-text transcribes. An LLM answers in text. Text-to-speech reads the answer out. In the middle sits a small piece that sets the rhythm of the whole conversation: turn detection.

export function createTurnDetector(onTurnEnd: () => void, silenceMs = 700) {
  let silenceStartedAt: number | null = null;
 
  return (isSpeech: boolean, now: number) => {
    if (isSpeech) {
      silenceStartedAt = null;
      return;
    }
    silenceStartedAt ??= now;
    if (now - silenceStartedAt >= silenceMs) {
      silenceStartedAt = null;
      onTurnEnd();
    }
  };
}

That is the core, simplified. A VAD says whether a frame contains speech. Once the silence crosses the threshold, the user is "done" and the pipeline fires. You pick the number, and whatever you pick is wrong for someone. Too low, and the agent talks over anyone who pauses to think. Too high, and every reply lands after a gap that sounds like a dropped call. Even the big providers' speech-to-speech APIs, like OpenAI's Realtime, have you configure turn detection, silence_duration_ms included. The model is native audio. The turn is still decided by a separate component.

The second problem is what survives the trip through text. A frustrated sigh becomes nothing. Sarcasm becomes the literal sentence. A voice cracking with sadness becomes the same string as one said calmly. Block puts exactly this at the center of the launch: a voice assistant that cannot hear you are in pain and just keeps going.

Full-duplex is not new. Kyutai's Moshi already listened and spoke on two parallel streams in 2024. But it sat on a 7B text LLM and generated an "inner monologue" in text ahead of the audio. What Block claims is that Dessa has no text anywhere.

What changes when the model owns the turn

If the model listens while it talks, the silence threshold has no job. There is no moment where the user is "done". There is a continuous stream both ways, and the model learned from real conversations when to jump in and when to let it run. That is what you hear in the demos, and it is why they do not sound like an agent.

What I do not know, and Block does not say, is how you wire a model like this into the rest of a system. A real voice agent has to check a booking, look up an order, open a ticket. In a cascade, text is the natural hook: you pass it to a function and get a result back. Where does a model that has never seen text read the order number from? I don't know, and I would rather say so than invent an architecture.

So today Dessa does not change your stack. It changes where you should invest less: that function with silenceMs in the middle. Keep it behind an interface you can swap, because it is the first thing a full-duplex model makes obsolete. It is the same call I make when deciding where a model is needed and where an if is enough. Except here, the if is the very thing the model does better.

"Indistinguishable" is also a problem

Block did not open Dessa to everyone, and after listening I understand why. Early access went to AI safety researchers with an explicit brief: find the risks the team has not anticipated.

The obvious one is the phone. Today you spot an automated caller within two seconds. With a voice like this you don't. The quality that makes an assistant pleasant to talk to is the same quality that makes a scam call believable.

When I will open that page again

The next time someone asks me for a voice assistant, before the pipeline gets locked in. I will play two Dessa demos and then the prototype, back to back, to remind myself where the bar is. Then I will write turn detection as a module I already know I will throw away.

The waitlist is open in the meantime. I have one question I want answered: when an API ships, will there be a way to call a function?