How Does Dessa Sound So Human?

That second voice is not a person. It is an AI called Dessa, and it learned to talk the way you did: just by listening.

How is Dessa different from most voice assistants?

Most voice assistants work like a walkie-talkie. You talk, it waits, and only when you stop does it answer. That delay happens because it turns your voice into text, sends the text to an AI to write a reply, and then turns that reply back into a voice.

Dessa works like a phone call. Engineers call this full-duplex, which means it listens and talks at the same time. So it can say "mm-hmm", laugh with you, or stop when you cut in.

What happens inside?

Dessa keeps two tracks running side by side: yours and its own. Sound is chopped into slices of about eighty milliseconds each. For every slice, the AI picks one of four moves: stay quiet and keep listening, say "mm-hmm", start talking, or stop because you just cut in. That is twelve decisions a second, nonstop.

It also has to tell sounds apart. "Mm-hmm" means keep going, while "Wait, no" means stop. Nobody writes these rules down. It learns them from thousands of real conversations.

Is this new?

Not exactly. Researchers have built duplex voice models for years. Kyutai released an open one called Moshi in 2024, and Google's Gemini Live and OpenAI's new ChatGPT voice do it too. Those started with an AI that learned language by reading text, and then had a voice added on top. Dessa was never trained on text. It was trained from scratch on recordings of human conversation and nothing else.

Is Dessa smarter than ChatGPT's voice?

No. It sounds more human, but sounding human and being smart are different skills. Dessa has no big brain behind it, at least not yet. Everything it knows, it heard during training.

The big players solve this differently. When you talk to ChatGPT's voice, GPT-5.5 does the heavy thinking in the background. Gemini Live can look things up and use tools while it keeps talking. ElevenLabs' new Eleven Realtime models talk and listen at once, but still keep text in the middle, connected to big AI brains, because business customers need agents that are intelligent, use tools and keep a record of every word.

Its CEO put it simply: all-in-one models are quicker and more expressive, but less reliable. ElevenLabs chose reliable. Dessa chose human.

Can you try Dessa?

Not yet. It is a research preview with a waitlist, and Block, its parent company, is letting AI safety researchers test it first.

What to remember

Think of Dessa as a child that learned to speak only by listening: completely natural, with a small brain. The big assistants are like an expert who only reads and writes: brilliant, but never fully natural. The race now is to build both in one, a voice that sounds like a person with a mind like an expert.