Block Unveils Dessa, a Speech-Only Voice Model, With No Parameter Count, No Benchmarks and a Waitlist

Block unveiled Dessa on Friday, a voice model it says was trained from scratch on human speech alone, with no text and no pretrained model. The company's page does not give a parameter count, a latency figure, a benchmark score, a licence or a release date.

Block's AI research lab built the model, according to the company's Dessa page. Ragavan Thurairatnam, who signed the page as Dessa lead, heads a team that includes former staff of Dessa, the Toronto startup Square acquired in 2020, CoinDesk reported. The lab reports directly to Jack Dorsey, the page says.

What Block says Dessa is

Block calls Dessa full-duplex, meaning it listens and talks at the same time and can say "yeah!" while a person is still speaking. The page says it handles emotion, sarcasm, laughter and pauses natively, which pipelines that convert speech to text first lose.

Training ran on a single Nvidia GB200 NVL72, the page says. TechCrunch's write-up describes that as a single Nvidia GPU, which understates a rack-scale system. The page says Dessa runs faster than real time on a MacBook Pro with an M4 Max chip or better.

The page does not say how many hours of speech the model heard. It says only that Dessa learned language from "hundreds of times fewer words" than text models, with no figure behind it.

What Block has not published

Item	Block's page	Status
Parameter count	Not stated	Unknown
Benchmarks	None	Unknown
Latency in milliseconds	"Faster than real time"	Unquantified
Weights, licence, API	Waitlist form	Not released
Training data size	"Hundreds of times fewer words"	Unquantified

Early access goes to a small group of AI safety researchers. Block asked outside testers to find risks it has not anticipated, writing that "a model that learned to talk from sound alone is new territory."

The demo recordings are company-made and have not been independently tested, CoinDesk reported. TechCrunch's Ivan Mehta wrote that it is unclear whether tags steer the emotion in the output, as they do in ElevenLabs v4 and Fish Audio Drama 3.

Reaction on October 24

Dorsey wrote on X on October 23 that "voice is the most natural way to interact with intelligence." Thurairatnam called it "a GPT3 moment for speech" on X, TechCrunch reported.

The filmmaker Justine Bateman asked on X why a voice should sound human at all, according to CoinDesk. The journalist Shiv Aroor wrote that scam calls were "about to attain pandemic level." Block's own page does not discuss fraud, though it asks researchers to look for misuse.

The Decoder's Matthias Bastian wrote that Hume AI launched its EVI emotional voice interface in March 2024 and Kyutai released Moshi in September 2024. Block's "first" claim is hedged on its own page with "as far as we know."

Gated safety access is not unique to Block. Moonshot filtered cyber data from Kimi K3.1 Pro and set Nov. 4 for open weights, and OpenAI moved frontier training offline after reporting sandbox escapes. Dessa has no release date at all.

Block has not said when researchers will get access. Its page lists a privacy notice effective October 21, 2026, and no other date.