Block Unveils Dessa, a Voice AI Trained Only on Human Speech

Illustration of two overlapping speech-bubble waveforms conversing, representing Block's full-duplex Dessa voice AI model

Block Unveils Dessa, a Voice AI Trained Only on Human Speech

Block, the payments company led by cofounder Jack Dorsey, has unveiled Dessa, an experimental speech-to-speech AI model that its research lab says was trained entirely on human conversation audio, with no text and no pretrained language model involved. The model can listen and talk at the same time and runs locally on a consumer Mac, according to Block’s own announcement.

Last updated: October 25, 2026, 05:40 IST.

Illustration of two overlapping speech-bubble waveforms conversing, representing Block's full-duplex Dessa voice AI model
Illustration: Praman News (AI-generated, editorial illustration)

What Block announced

According to Block’s own announcement page, Dessa is a “full-duplex” speech model, meaning it can hear and generate speech at the same time rather than waiting for a pause before responding, similar in concept to turn-taking in a real conversation. The project comes from a small AI research lab inside Block that reports to Dorsey, Block’s cofounder and chairman, and was founded by former members of Dessa, a Toronto deep learning startup the model is named after.

Block says it trained Dessa “from scratch on nothing but human conversation,” without text data or an existing pretrained language model feeding it knowledge. The company states the model learned from audio alone using what it describes as “hundreds of times fewer words” than large language models typically require, though that comparison is Block’s own characterization rather than an independently verified benchmark.

How it was trained and how it runs

Block says Dessa was trained on a single NVIDIA GB200 NVL72 system, a high-end AI training rack, a notably small training footprint compared with the large multi-rack clusters typically used to train frontier language models. For inference, Block says the model runs locally on a MacBook Pro with an M4 Max chip or better, and does so faster than real time, meaning it does not need a cloud server to hold a conversation once deployed on suitable hardware.

Ragavan Thurairatnam, who leads the Dessa project at Block, is quoted on the announcement page as saying the team believes Dessa is “the world’s first AI you can have a realistic live conversation with that was trained only from human speech.”

What the model can do

Block’s page describes Dessa as capable of producing emotional speech, including sarcasm, laughter and sighs, and of using natural pauses and cadence rather than flat, machine-paced delivery. The model is also described as knowing when to interrupt and when to offer brief backchannel responses, such as “yeah” or “got it,” the small verbal cues people use to show they are listening without taking over the conversation.

Block says some behaviors shown in its demo recordings, such as the model getting louder when asked to speak up, were not explicitly programmed or trained for, suggesting they emerged from the speech-only training approach rather than being hand-coded. The company has also published demos pairing Dessa with a third-party conversational model called GPT-Live-1, though Block notes that GPT-Live-1’s developer is not affiliated with Block and has not reviewed or endorsed Dessa.

Who can try it, and when

Dessa is not yet publicly available. Block says it is first giving early access to a small group of AI safety researchers, who must apply with an institutional email address and, optionally, a Google Scholar profile or blog link. Block’s page says general “launch links are coming soon” but gives no public release date.

Limiting early access to safety researchers suggests Block is treating the model’s realism, including its ability to mimic natural human conversational behavior, as something that needs scrutiny before a wider release, a caution that has become more common across the AI voice industry as cloned and synthetic voices have been used for scams and impersonation.

Why a payments company is building a voice model

Block is best known as the parent of Square and Cash App, not as an AI research company, which makes Dessa an unusual release for it. Square, Block’s former name, announced on February 7, 2020 that it had acquired the Toronto machine learning startup Dessa, known at the time for deepfake-detection research, in a deal with undisclosed terms. TechCrunch reported that former members of that team now operate as a small, semi-independent research lab inside the larger company and named the new voice model after it.

Dessa enters a voice AI field that already includes OpenAI’s GPT-Live effort, which Block’s own demo page uses as a pairing partner for some of its sample conversations. Full-duplex, emotionally expressive voice models have become a competitive focus for AI labs over the past year, as companies look past text chatbots toward more natural spoken interfaces for phone-style assistants, customer service bots and accessibility tools.

Frequently asked questions

Can I use Dessa right now? No. Block has only opened a waitlist and is currently granting access to AI safety researchers, not the general public.

Does Dessa need an internet connection to work? Block says it runs locally on a sufficiently powerful Mac, which suggests conversations can happen without sending audio to a cloud server, though Block has not detailed the exact deployment architecture for any eventual public release.

Source: Block.