Block Dessa AI: speech-only voice model runs on a MacBook Pro

Key takeaways

• Block, the company led by Jack Dorsey, has unveiled Dessa, a full-duplex speech-to-speech voice AI model that listens and talks at the same time.
• Block says Dessa was trained from scratch only on recordings of human conversation, with no text and no pretrained models, on a single Nvidia GB200 NVL72 rack.
• The company says the model runs faster than real time on a MacBook Pro with an M4 Max chip or better.
• Dessa is a research preview. Only a small group of AI safety researchers has access, and there is no public release date.
• Block has not published the model's size, benchmarks, supported languages or training-data volume. Critics on X warned that a voice this human could make scam calls more convincing.

What is Block's Dessa AI?

Block Dessa AI is a full-duplex speech-to-speech voice model that listens and talks at the same time. Jack Dorsey announced it on X on October 23, 2026 (18:23 UTC), and Block says it was trained only on recordings of human conversation, with no text.

Dorsey wrote that the team sees voice as the most natural way to use AI, and that talking with Dessa "feels like talking with a person." Block's project page goes further and calls Dessa "the most authentic full-duplex speech model ever built". That is the company's own claim, and no outside test has checked it yet.

The model comes from a small AI research lab inside Block that reports directly to Dorsey. Ragavan Thurairatnam leads the lab and co-founded the Toronto startup Dessa, which Square, now Block, bought in February 2020. According to TechCrunch, Thurairatnam called the release a "GPT3 moment for speech" in a post on X.

How Dessa works: full-duplex, speech only

Most voice assistants work like a relay. Speech is turned into text, a large language model writes a reply, and a text-to-speech engine reads it aloud. Each step adds delay, and the text step drops information such as tone, hesitation and laughter. In a blog post quoted by TechCrunch, Block said current voice AI "can just plain suck" for this reason.

Block describes Dessa as different in three ways:

• Full-duplex. The model listens while it speaks, so it can be interrupted, can interrupt, and can give short backchannel replies such as "yeah" or "got it" while the other person is still talking.
• Speech-only training. Block says Dessa was trained from scratch on the sounds of human conversation, with no text and no pretrained models. The company says it learned language from "hundreds of times fewer words" than text-based LLMs, but it gives no hours of audio or dataset size.
• No separate LLM. According to TechCrunch, Block says Dessa's knowledge comes from the conversational data itself, not from a paired language model that speech systems usually rely on for facts.

Block says the model hears and produces emotion, sarcasm, laughter, sighs and voice cracks, and that it picks its own voice and personality. TechCrunch noted it is not clear whether that emotion can be steered, as it can with the voice tags used by tools such as ElevenLabs v4 and Fish Audio Drama 3. The demo clips on Block's page are live recordings of people talking to Dessa, with subtitles added later. CoinDesk pointed out that these are company demonstrations, not independent tests.

Dessa at a glance: what Block has and has not disclosed

Information as of October 24, 2026, 14:00 UTC.

Trained on one Nvidia GB200 NVL72

The training claim is one reason the launch drew attention. Block says Dessa was trained on a single Nvidia GB200 NVL72. That is a large machine, but it is one rack, not the data-centre clusters behind leading text models. According to Nvidia's product page, one GB200 NVL72 contains:

Block has not said how long training took, so the total compute cannot be estimated from public information.

Can Dessa run locally on a MacBook?

Block says yes: inference runs faster than real time on a MacBook Pro with an M4 Max chip or better. TechCrunch also reported that Thurairatnam described the model as runnable on consumer GPUs. A voice model that fits on a laptop would not need to send audio to a cloud server, which matters for privacy and cost.

Without a parameter count, outsiders cannot yet check how much memory Dessa needs or whether it has been compressed. For background on how large models are shrunk to fit consumer hardware, see our explainer on LLM quantization and GGUF. Our guide on how to run a local LLM with Ollama covers how text models run on a laptop today. Dessa itself is not available to download.

Who can use Dessa now?

Almost no one. Block's page offers a waitlist and says a small group of AI safety researchers is getting early access first. Researchers can apply with an institutional email address and, optionally, a Google Scholar profile or blog. Launch links are marked "coming soon", and CoinDesk reported that no public release date has been given.

According to TechCrunch, Block asked those researchers to focus on whether a voice agent could deceive people. On its own page, Block says a model that learns only from speech is new territory and that it wants outside researchers to find risks the team did not expect.

Voice-cloning and scam concerns

Reaction on X split quickly. Some users praised the demos. One post quoted by CoinDesk said its author had never heard anything artificial that sounded so human. Binance co-founder Changpeng Zhao wrote that it "might be very good" for teaching children, a task Zhao said current AI voices make hard.

Others saw a risk. Filmmaker Justine Bateman asked why AI voices need to sound human at all and wrote: "So you can trick people into lowering their guard." Journalist Shiv Aroor posted that "scam calls about to attain pandemic level," according to CoinDesk.

Block has not said whether Dessa can copy a specific person's voice. Its page says the model chooses its own voice and that no one steers the demo recordings. CoinDesk noted that the debate so far is about how human the model sounds, not about cloning. The concern still has history behind it. Before Square bought the original Dessa in 2020, the startup was best known for deepfake detection, and VentureBeat reported that its technology was behind a convincing fake audio clip of podcaster Joe Rogan. When Square announced the deal, it cited the rising risk of deepfake-related fraud as one reason for buying the company.

Block also argues that hearing emotion is a safety feature. Its page says today's voice AI is nearly deaf to how people feel, and gives the example of a user in serious pain whom the system fails to notice. A speech-trained model could, in Block's view, pick up distress that a transcript misses. No independent study of that claim exists yet.

Why Block, a payments company, built a voice model

Block runs Square, Cash App and the Bitkey bitcoin wallet. TechCrunch noted that Square already uses voice AI for ordering, and suggested that a model running on a device could make more business and consumer uses cheaper. That is the reporter's reading; Block has not announced any product plans for Dessa. Block says it built the model out of curiosity, to improve understanding between humans and AI, and for safety.

The launch also puts Block in a crowded field. TechCrunch compared Dessa with OpenAI's GPT-Live, another full-duplex model released earlier in 2026. One of Block's demo clips shows Dessa talking to GPT-Live-1. Block notes that GPT-Live-1's developer is not affiliated with Block and has not endorsed Dessa. Voice is a natural interface for the tool-using assistants described in our tutorial on building an AI agent with tool calling, so the contest is likely to grow.

What it means

If Block's claims hold up, Dessa shows that a natural-sounding conversational voice model can be trained on speech alone, on one rack of GPUs, and run on a laptop. That would make it cheaper to build voice agents that do not rely on a cloud LLM.

The open questions are large, though. Block has published no model size, no benchmarks, no languages, no training-data details and no release date, so outside experts cannot compare Dessa with GPT-Live or ElevenLabs yet. The scam concern is also real: the more a machine sounds like a person on the phone, the less a caller's voice proves who they are. The safety researchers' findings, and whether Block publishes them, will show whether Dessa becomes a product or stays a research demo.

Sources

• Block: Dessa, the conversation model trained only on speech
• Jack Dorsey on X: announcement of Dessa (October 23, 2026)
• TechCrunch: Block's new voice model sounds eerily human
• CoinDesk: Block's new voice AI sounds eerily human, sparking awe and scam fears
• Nvidia: GB200 NVL72 specifications
• Square press release: Welcoming the Dessa team to Square (February 7, 2020)
• VentureBeat: Square acquires deepfakes research firm Dessa