By ByteBot
Block unveiled Dessa on October 23 — a full-duplex AI that learned conversation the way humans do: by listening to other humans talk, not by reading what they wrote. No text. No pretrained models. The Block Dessa speech model runs audio in, audio out, trained on a single rack of 72 NVIDIA GPUs. It runs locally on a MacBook Pro M4 Max and responds faster than real time. Block calls it “a GPT-3 moment for speech.” That’s a bold claim, and for once, the technical story behind it is genuinely interesting.
What Speech-Only Training Actually Means
Every major voice AI model in use today handles speech through text as an intermediary. The standard pipeline converts audio to text, runs a language model over that text, and converts the response back to audio — three steps, three places where paralinguistic information gets dropped.
The moment you transcribe audio to text, you lose tone, hesitation, sarcasm, and the way a voice cracks when someone is upset. A sentence like “sure, that sounds great” means something different said with resignation than with enthusiasm. Text cannot carry that distinction. Language models trained on text have never heard the difference.
Dessa skips the whole pipeline. Audio goes in, audio comes out. Block says the model trained on human conversation recordings and nothing else — no text corpus, no pretrained weights from a language model underneath. As the Block team framed it: “Imagine being in serious pain while a voice AI has no idea.” That is the gap they are targeting. This is not a pipeline optimization. It is a different bet about what the right foundation for voice AI looks like.
Full-Duplex Done Differently
Most voice AI still runs half-duplex: you talk, it waits, then it responds. Full-duplex changes that entirely. Dessa listens and speaks simultaneously, which means it can be interrupted mid-sentence, interject “got it” while you are still talking, and adjust its timing to yours — the phone-call model rather than the walkie-talkie model.
OpenAI’s GPT-Live-1 is also full-duplex. However, the architecture is different: GPT-Live-1 delegates complex reasoning to a backend text model. When you ask it something that requires structured thinking, a language model runs behind the scenes doing the heavy lifting in text. Dessa has no backend text model at all. What you hear is what the model figured out from audio alone.
That is genuinely fascinating — and it is also where the product gaps start to show. No one outside Block has verified what Dessa can and cannot handle when a conversation moves beyond small talk into something that requires memory, multi-step reasoning, or structured output. Block has not said whether tool calling is on the roadmap. The demos are impressive. The benchmarks do not exist yet.
Why Local Inference Matters More Than You Think
The fact that Dessa runs on a MacBook Pro M4 Max matters for more than latency bragging rights. Voice AI that runs locally means conversation audio never leaves the device. For developers building voice-enabled applications in healthcare, legal, financial services, or enterprise environments where audio data has compliance implications, on-device inference is not a nice-to-have. It is a hard requirement.
Furthermore, Block trained the model on a single NVIDIA GB200 NVL72 rack — 72 GPUs — which is lean compared to frontier text models. They also claim it required “hundreds of times fewer words” to train than a comparable language model. If that efficiency scales, speech-native models could be meaningfully cheaper to develop and deploy than text-grounded counterparts.
What Block Has Not Told Developers
Block has released a demo and a waitlist. What it has not released: parameter count, dataset size, latency benchmarks, accuracy comparisons against GPT-Live-1, or any indication of when weights might be available. The current access path is a public waitlist — institutional email required — and early access for a small group of AI safety researchers specifically tasked with finding risks “Block hasn’t anticipated.”
That is not a product launch. That is an honest research preview. Releasing access to safety researchers before the general developer community is the right call for a model that touches live audio and emotional inference. It also means developers cannot build with Dessa today — and should not treat this as a near-term vendor decision.
The open question every voice agent developer should sit with: can a model that has never processed a sentence in text form handle the reasoning demands of a real-world voice application? Speech-native training is the right long-term direction. Dessa as announced today is a proof of concept with significant reach. The benchmark tests — when they arrive from independent researchers — will determine how far that reach extends.
How to Get Access
The public waitlist for Dessa is live at Block’s research page. Institutional email is required. No general release date has been announced, and Block has not indicated whether model weights will be made publicly available. For broader context on Dessa’s full-duplex design and training approach, Unite.AI has the most complete technical write-up available. Meanwhile, CryptoBriefing covers the broader Block context — including why Jack Dorsey’s payments company is funding speech AI research in the first place.
Watch the early-access safety research results. Those will tell you whether Dessa is worth integrating into a voice pipeline or whether it remains a technically important research artifact waiting for its benchmark moment.
