Deep Dive
How Dessa AI Works: Full-Duplex Speech-to-Speech, With No Text in the Middle
To understand why Dessa matters, start with how the voice AI on your phone works today — because Dessa is a rejection of almost every part of it.
The old pipeline: speech → text → speech
Traditional voice assistants are a relay race with three runners. Automatic speech recognition transcribes what you said into text. A large language model — trained on trillions of written words — generates a text reply. Then text-to-speech reads that reply aloud. The system works, but each handoff bleeds information. Your sarcasm, your hesitation, the crack in your voice, the laugh in the middle of a sentence: none of it survives transcription. The words arrive; the meaning half-leaves. Block's verdict is blunt: “Interacting with current voice AI can just plain suck… The most powerful AIs in the world can't hear how you feel.”
Dessa's bet: delete the middle step
Dessa takes audio in and produces audio out, directly. Everything it knows came from human speech — Block trained it from scratch on conversation recordings, with no text data and no pretrained language model underneath. That is a much harder learning problem than it sounds. A single spoken word takes thousands of times more bytes than the same word in text, and speech never repeats exactly: say the same sentence twice and the data differs both times. Text models get clean, identical tokens. Dessa got the messy real thing — and had to extract language, facts, timing and emotion from it simultaneously.
What “full-duplex” really means
Most voice AI is half-duplex, like a walkie-talkie: strict turns, waiting for silence before replying. Full-duplex is a phone call. Dessa processes incoming speech while producing outgoing speech, so it can react during your sentence — backchanneling (“yeah!”, “got it!”), laughing at the funny part rather than after it, noticing an interruption and yielding, or choosing to cut in itself. Timing stops being a bug and becomes part of the conversation. That is why Block's demos feel different even when the words are ordinary: the when is human, not just the what.
Emotion is native, not tagged on
Competing expressive voice systems often fake feeling with written cues — tags that instruct a voice to laugh or whisper at a marked point. Dessa's emotion is not an annotation layer. Because it learned from real conversations, Block says it natively hears and produces sarcasm, laughter, frustration and sadness, adjusts its volume when asked, and shifts intonation and speed across a wide range. The safety angle matters here too: Block argues a voice AI that cannot detect distress — real pain, real panic — is not just a worse product, it's a potentially dangerous one.
The efficiency story: one rack to train, one laptop to run
Two numbers define Dessa's engineering story. Training: a single NVIDIA GB200 NVL72 — one rack-scale system — rather than a sprawling GPU cluster, and a training diet of hundreds of times fewer words than LLMs consume. Inference: faster than real time, running locally on a MacBook Pro with an M4 Max chip or better. Local inference is more than a party trick. It hints at voice AI with no per-minute cloud bill and audio that never has to leave the device — economics that would matter enormously in, say, the restaurants and shops already running on Block's Square platform. That last part is informed speculation about where this could go; Block itself has announced no product, no API and no price.
The honest caveats
Everything above traces to Block's own claims and its hand-picked demo recordings. Block has published no parameter count, no training-data hours and no benchmark scores, and no independent researcher has yet published a hands-on evaluation. A model that learned only from conversation will know a lot about how people talk and — by design — has never read a book, a manual or a webpage. Whether speech-only learning scales to precise, factual, tool-using assistant work is the open research question Dessa now poses to the whole field. Thurairatnam has called the launch “a GPT-3 moment for speech.” The next year of independent testing will decide whether that label sticks.
Sources: Block's official Dessa page; Kyutai's Moshi paper (arXiv:2410.00037) for the full-duplex comparison; TechCrunch and The Verge, October 2026.
Frequently asked questions
What does full-duplex mean for Dessa AI?
It means Dessa listens and speaks at the same time, like a person on a phone call. It can backchannel, laugh mid-sentence, handle interruptions and react in real time instead of waiting for you to finish and fall silent.
Was Dessa trained on text?
No. According to Block, Dessa was trained from scratch on human conversation audio only — no text data and no pretrained models of any kind.
What hardware does Dessa need to run?
Block says inference runs faster than real time locally on a MacBook Pro with an M4 Max chip or better. Training used a single NVIDIA GB200 NVL72 system.
🚀 Get Dessa launch alerts
Dessa is still a research preview. Join the free DessaAI Daily list and we will email you the day access opens — plus the week's biggest Dessa stories, once a week. No spam, no hype.
Form not loading? Subscribe on the signup page →