Dessa AI Explained: Block’s Voice Model That Learned to Talk Without Ever Reading a Word

Dessa AI is a new voice model from Block, the company behind Square and Cash App, and it is unusual for a talking machine because it never learned to read. Block says it trained the model only on recorded human conversation, with no text at all. Block also says Dessa AI listens and speaks at the same time, the way a person does on the phone. This post is for creators, podcasters and developers who follow AI audio. It covers what Block showed on October 23, how the model differs from the voice assistants you already use, where the name comes from and what nobody outside Block can check yet.

The demos are good. In places they are unsettlingly good. One of them gasps.

But Block recorded every clip and picked every clip, and nobody has published a benchmark yet. So read everything below as “Block says” unless a link points somewhere else.

Dessa AI is Block’s voice model that skips text entirely

Block launched Dessa AI on its own Dessa page with a waitlist, a stack of recorded conversations and a call for safety researchers. TechCrunch and WinBuzzer covered it over the next two days. The pitch is short. Audio in, audio out, nothing written down in between.

The lab behind it sits inside Block. Block calls it “(very) small”, and the brackets are Block’s. According to WinBuzzer, Ragavan Thurairatnam leads it, and the lab reports directly to Jack Dorsey, Block’s cofounder and chairman, which is an unusually short line for a team that says it has built deep learning models since 2012 and generative speech models for more than eight years.

Three claims carry the announcement. The model is speech-to-speech, so it hears sound and produces sound. It is also full-duplex, which means it can talk while you talk. And everything it knows, from grammar to movie titles, came from listening. Block calls Dessa AI, “as far as we know”, the first AI you can have a realistic live conversation with that was trained only on human speech.

Nobody outside can try Dessa AI, though. There is a waitlist. That is all the public gets for now.

Most voice assistants pass your words through three models

Most voice assistants are three systems taped together. Speech recognition turns your voice into text, then a text model (an LLM, the kind of model behind chatbots) writes an answer, and finally text-to-speech reads that answer aloud. Every handoff costs time. Every handoff loses something.

What gets lost is everything that isn’t a word. Say “great” right after your laptop dies, and the transcript says “great”. The sarcasm stays behind. Block’s announcement makes exactly this complaint: current voice AI “throws away an incredible amount of information text alone can’t carry.”

A speech-to-speech model gets rid of the handoffs. Sound goes in, sound comes out, and the model works on the audio itself, so tone, speed, a sigh and a laugh all count as input.

Full-duplex is the other half. A walkie-talkie is half-duplex: one person talks and lets go of the button, then the other answers. A phone call is full-duplex, because both people can speak at once, interrupt and overlap. Most assistants still behave like walkie-talkies. They wait for silence.

A backchannel is the small noise a listener makes so the speaker knows they are still there: “mm-hm”, “yeah”, “ooh”. Nobody writes these down. Still, a conversation without them feels like leaving a voicemail. Block says Dessa AI learned when to backchannel and when to cut in from real conversations, not from a rule someone wrote.

What “trained only on speech” means for Dessa AI

Large language models learn from text and need an enormous amount of it, while Block says Dessa AI learned language from “hundreds of times fewer words” than an LLM needs. In fairness, the company adds, LLMs also have to learn C++ “and whatever the hell Angular is.”

(Angular is a web framework. We have opinions about it too, mostly formed at 2 a.m. while a build failed for reasons only the build understood. Anyway.)

Fewer words doesn’t make the job easier. Text is tidy, because “stole” is the same five letters whoever types it. Speech never is. Say one sentence twice, Block notes, and the two recordings still come out different. A spoken word also takes thousands of times more bytes than the written one, so the model has to find the language inside a lot of noise.

Block compares Dessa AI to babies, who learn to talk before they can read. It also says it trained the model on a single NVIDIA GB200 NVL72, which WinBuzzer describes as one rack holding 72 GPUs. That is a lot of hardware for a lab that calls itself very small. Still, it is one rack. Not a data centre.

Stress one word and the sentence changes meaning

Block’s own example explains the point best. Read “I didn’t say she stole my CUDA code” aloud and stress a different word each time. Stress “I” and someone else said it, put the weight on “she” and the thief is someone else, lean on “my” and she stole somebody else’s code.

We tried this around a desk. It works. Afterwards everybody looked a little guilty.

That is the whole argument. A text pipeline sees the same eight words every time, while a speech model hears several different accusations, and only one of them gets someone fired.

What Block doesn’t say is where the conversations in the training set came from or how it got permission to use them. WinBuzzer pointed out the same gap.

What the Dessa AI recordings show

Block’s page has about twenty clips. Each one is a live recording of a person talking to the model, with two speaker tracks and subtitles added afterwards. Block stresses that nobody steers Dessa AI, so it picks its own words, timing, voice and personality.

Here are the highlights, as Block describes them. It laughs mid-sentence when something is funny instead of waiting for you to finish. Ask it to speak up. It gets louder. Blank on a movie title and it finishes the sentence for you. In the anniversary clip it gasps “oh my god” under its breath, then answers with what Block calls teary gratitude. TechCrunch also describes a clip where the bot says it is tired from work.

It sounds tired.

The strangest clip has no humans in it. Block put Dessa AI on a call with GPT-Live-1, which its page calls ChatGPT Live, so both voices are AI-generated. Block adds a disclaimer that the other model’s developer has not reviewed or endorsed the demo. According to TechCrunch, GPT-Live is a full-duplex model OpenAI released earlier this year.

Dessa AI makes mistakes, and Block kept them in

The mistakes stayed in. As WinBuzzer reports, The Verge noticed Dessa AI calling its partner Ronaldo. In another clip, a chat about Christopher Nolan’s films drifts into Bryan Singer’s Superman films. Thurairatnam also shared a guessing game played with a Block intern, and Gizmodo noted that the model guessed right on the first try twice, then kept guessing after it was told it had won.

Now for the caveat. Block chose these recordings. There is no technical paper, no benchmark and no public test. TechCrunch also says it is unclear whether some kind of tag steers the emotion, the way ElevenLabs v4 or Fish Audio Drama 3 use voice tags. “Nobody steers it” is a claim, and for now nobody outside Block can check it.

Moshi and PersonaPlex did full-duplex first, with text underneath

Talking and listening at once is not new. In 2024 the lab Kyutai published Moshi, which it calls a speech-text foundation model and full-duplex spoken dialogue framework, and which the paper describes as the first real-time full-duplex spoken large language model. The reported delay is 160 milliseconds in theory. About 200 in practice.

The difference is in the phrase “speech-text”. Moshi predicts text tokens that line up with its own speech, a trick the authors call Inner Monologue. It is also built on a text language model called Helium. Text, in other words, holds Moshi’s speech up from underneath. NVIDIA’s PersonaPlex builds on Moshi’s architecture and adds a way to set the model’s role with a text prompt and its voice with an audio sample.

So Block’s claim is narrower than “full-duplex”, and stranger. Block says Dessa AI drops the text support entirely, with no text model underneath, no transcript and no pretrained weights, and if that survives outside testing, it answers a research question more than a product one. Can knowledge and conversation both come from sound alone?

Block lists more open questions itself. How efficiently do these models learn facts compared with text models? How quickly can they pick up a new language? Good questions. Answering them will take papers, not demos.

Running on a MacBook Pro changes who could use Dessa AI

Block says the trained Dessa AI model runs locally on a MacBook Pro with an M4 Max chip or better, and that it produces audio faster than real time, so there is no awkward pause before it answers. For comparison, WinBuzzer reports that Moshi’s maintainers tested their on-device version on a MacBook Pro M3.

Local matters. The audio never has to leave the machine. Nobody pays a server bill by the minute. And the response time doesn’t depend on someone else’s data centre having a good day.

TechCrunch reads this as a business move. Square already uses voice AI to take orders, and a model that runs on the device could power more business and consumer products at a lower cost. That is TechCrunch’s analysis, though, not Block’s roadmap. Block has announced no product, no API and no price.

The Dessa name comes from the fake Joe Rogan voice

Now the plot. Dessa was a Toronto deep learning startup that started out in 2016 as DeepLearni.ng, according to BetaKit. In May 2019 its engineers released clips of Joe Rogan saying things Joe Rogan never said. Convincing, and entirely fake. The system was called RealTalk, and The Register reported that Dessa would not release the research, the model or the datasets.

RealTalk was text-to-speech: it took typed words and made them sound like one particular man. The Logic notes that Dessa also open-sourced tools to help others detect deepfakes.

Then a buyer showed up. In February 2020 Square bought Dessa for an undisclosed sum, and Square’s own announcement leaned on the deepfake work, calling deepfake audio and video a growing security, risk management and fraud threat. Square renamed itself Block in 2021. When the deal closed, Thurairatnam was Dessa’s head of machine learning, per BetaKit. Now he leads the lab whose model carries the old company’s name.

So the same people who once showed how convincing a fake voice could be, made from text, have now built a talking voice that has never seen text. Somewhere a 2019 risk memo is reading this with great interest.

The deepfake question changes shape

A voice that gasps, laughs on time and sounds tired when it says it is tired would help anyone building a patient assistant. It would help a scammer too. Block seems to know it. Instead of a public launch, it is giving early access to a small group of AI safety researchers so they can find risks the team didn’t anticipate. TechCrunch suspects one of those tests is a voice agent tricking the person on the other end of the line.

Block’s own safety argument points the other way. Its page asks you to imagine being in serious pain while a voice AI has no idea and just keeps going. A model that hears distress could stop, and Block hopes Dessa AI will, although as WinBuzzer points out, nobody has publicly tested whether it actually does, or how reliably.

Two details are worth watching. First, the model picks its own voice and personality for each conversation, which is charming in a demo and confusing on a phone call. Second, in the GPT-Live-1 clip Block labels both voices as AI-generated, and that label is the only way a listener knows. Future recordings, made by other people, may not have one.

What is still unknown about Dessa AI

Here is what Dessa AI still lacks in public, as of October 26:

• a release date, an API or a price
• benchmarks, a technical paper or any independent test
• a list of supported languages (one clip is titled “Singing nonsense Spanish”, which is not a language list)
• where the training conversations came from and whether the speakers agreed to it
• a parameter count, which Block’s page doesn’t give
• an answer to whether tags steer the emotion, the question TechCrunch raised

The dataset question will sound familiar to anyone who has followed AI music. Musicians have spent the past year asking whose recordings went into which model. Their union now wants a say before labels train on anything, as our post on the AFM’s “Human Made, Human Paid” campaign covers. Voices are recordings too.

What a speech-only model could change for audio makers

Opinion time.

For podcasters, the uncanny valley just got shallower. AI co-hosts have mostly sounded like audiobooks with good posture, and a model that backchannels and laughs over you wouldn’t. That makes disclosure part of the craft. Listeners forgive a lot, but they don’t forgive finding out late.

Editing changes too. Many podcast workflows now cut from the transcript, which works because the meaning lives in the words. In a conversation full of overlaps, half-laughs and a sigh in the middle of a sentence, the transcript misses the point. That is one reason we like editing on the sound itself, whether in a basic five-step podcast edit or in an audio-first tool such as Mubert Cast. The sigh never makes it into the text. So it gets edited where it lives.

For developers, the local angle of Dessa AI is the interesting part, but there is no API, no licence and no price, so there is nothing to build on yet. Nothing goes on a roadmap until a paper appears, then weights or terms.

For voice actors and anyone whose voice is their product, the open question is the training data. A model that learned timing, sarcasm and breathing from real conversations learned them from real people, and those people will want to know if they were in the training set.

Which brings back Joe Rogan. In 2019 the people behind this name made a fake Rogan read tongue twisters from a script. In 2026 their model gasps at an anniversary surprise, and there is no script at all.

A checklist for creators watching Dessa AI

1. Listen to three clips on Block’s Dessa page and note where the timing sounds human and where it slips.
2. Read the introduction of the Moshi paper to see how a full-duplex model with text underneath works.
3. Add a line about AI voices to your show-notes template now, before an episode needs it.
4. Agree with co-hosts and regular guests on how you will handle AI voices in an episode.
5. Check which voice tools in your workflow send audio to a server and which ones run locally.
6. Join Block’s waitlist if you want to hear the model live once access opens.
7. Treat Dessa AI demos as unverified until a paper or an independent benchmark comes out.

Cover: illustration based on a photo of Jack Dorsey.jpg) by Steve Jurvetson, CC BY 2.0.