Block, the company behind Square and Cash App, has shown Dessa, a voice model that learned to hold a conversation from recorded human speech alone. There was no text in its training data and no pretrained language model underneath it. It hears audio and answers in audio, and it can listen and talk at the same time.
That last part has a name, full-duplex, and it is why the demo clips sound so different from the voice assistants most people use. Dessa laughs halfway through your sentence, says “yeah” while you are still talking, and finishes your thought when you forget a movie title.
The launch is one of the most talked-about AI stories of the past two days. The Hacker News thread passed 500 points and 190 comments, the lead researcher’s announcement on X drew more than 3.4 million views, and X posts mentioning Dessa jumped from a few dozen a day to more than 7,500.
You cannot use it yet. Dessa is a research preview, and the only outside access is a small early program for AI safety researchers.
Key points
- Speech only: Block says Dessa was trained from scratch on human conversation audio, with no text and no pretrained model.
- Full-duplex: it listens and speaks at once, so it can interrupt, back-channel and react mid-sentence instead of waiting for its turn.
- Small training budget: Block says it trained Dessa on a single NVIDIA GB200 NVL72 system.
- Local inference: Block says it runs faster than real time on a MacBook Pro M4 Max or better.
- Not available: no API, no weights and no app. There is a waitlist and a safety-researcher access form.
What Block actually showed
Dessa comes from a small research lab inside Block, started by people from Dessa, the Toronto deep learning startup the model is named after. Block says the lab reports directly to Jack Dorsey and has worked on generative speech models for more than eight years.
The launch page is a set of live recordings of people talking to the model. Block says nobody steered the conversations and that Dessa picks what to say, when to say it and how, including its own voice and personality. Subtitles were added afterward.
The clips are built to show behaviour that text-based voice bots struggle with:
- It laughs the moment something is funny, while the person is still talking.
- When asked to speak up, it gets louder.
- When a speaker blanks on a movie title, Dessa fills it in.
- In one clip it reacts to an anniversary surprise with a gasp and what sounds like emotional thanks.
- In another it talks to GPT-Live-1, a third-party voice model, with both sides AI-generated.
Lead researcher Ragavan Thurairatnam called it “a GPT3 moment for speech” and said Dessa can run on consumer GPUs.
Full-duplex, explained simply
Most voice assistants work like a walkie-talkie. You speak, they wait for you to stop, they turn your words into text, a language model writes a reply, and a text-to-speech model reads it back. That chain is called a cascade. It works, but it has two built-in problems.
- Turn-taking: the system has to guess when you have finished. Guess too early and it cuts you off. Guess too late and there is an awkward pause.
- Lost signal: once speech becomes text, the tone, pace, sighs, laughter and sarcasm are gone. “Fine” said cheerfully and “fine” said through gritted teeth become the same word.
A speech-to-speech model skips the text step. Audio goes in and audio comes out, so tone can shape the answer. A full-duplex model goes further: it processes your voice and produces its own at the same time, like a phone call rather than a walkie-talkie. That is what lets it back-channel with “got it” or “ooh” and jump in when a person naturally would.
Dessa is not the first full-duplex model. What Block says is new is the training recipe: every word, fact and habit the model has came from raw sound, the way a child learns to talk before learning to read.
Why training on speech alone matters
Block makes two claims here that are worth separating.
The first is efficiency. Block says Dessa learned language from hundreds of times fewer words than a text LLM needs, and was trained on one GB200 NVL72 rack. If that holds up, high-quality voice models may not need frontier-lab budgets.
The second is safety. Block’s argument is that AI systems are gaining power while staying nearly deaf to emotion. A text pipeline cannot hear that someone is in pain or panicking. A model that hears tone natively could, in principle, react to it. That is why the first outside access is going to safety researchers rather than developers.
The catches
The reaction has been split, and the criticism is useful for anyone weighing what this means.
- Voice over substance: several Hacker News commenters said the delivery sounds human but the content is thin, with one calling the reasoning “2024-level”. Knowledge learned only from audio is likely narrower than what a text model absorbs.
- Interrupting a lot: some listeners found the constant cut-ins and filler sounds tiring rather than natural.
- Hand-picked clips: these are demos Block chose. Nobody outside the safety program has tested the model.
- Open questions: Block has not published model size, a paper, benchmarks, pricing or a release timeline. Developers asking for weights or an API have not had an answer yet.
- Misuse: a voice this convincing raises obvious worries about phone scams and emotional over-reliance, which commenters and Block’s own safety framing both point to.
What it means for builders
Nothing changes in your stack this week. There is nothing to call. But Dessa is a clear sign of where voice products are heading.
- Cascades are on borrowed time for conversational products. If you are building a voice agent today on speech-to-text, an LLM and text-to-speech, plan for a swap to a native speech-to-speech model where latency and tone matter.
- Turn-taking logic may disappear. A lot of voice-agent engineering goes into deciding when the user has stopped talking. Full-duplex models make that problem part of the model.
- On-device voice is getting real. If Block’s laptop claim holds, private, offline voice assistants become practical without a cloud round trip.
- Expect hybrids. The obvious next step, raised by many developers, is joint training on speech and text, so the model keeps Dessa’s timing and tone with an LLM’s knowledge.
What to watch
- Whether Block publishes a paper, model size or benchmarks.
- Whether the waitlist turns into an app, an API or open weights.
- What the safety researchers report back.
- How OpenAI, Google and others answer with their own full-duplex voice models.
Sources
- Block: Dessa, the conversation model trained only on speech
- Ragavan Thurairatnam’s announcement on X
- Block’s announcement on X
- Hacker News discussion
Image credit: Block
