JP — ITmedia

Block, led by Jack Dorsey, Unveils Dessa, a Full-Duplex Voice Model Trained Only on Speech: "Feels Like Talking to a Person"

AIOctober 26, 20261 min read

Block, the company led by Jack Dorsey, has shown a new voice AI called Dessa. It learns only from human speech, not text, and it can talk back in real time, like a person in a conversation.

What happened

On October 23 (US time), Block (formerly Square) announced Dessa as a research preview. Dessa is a speech-to-speech model. It takes voice in and gives voice out. It can listen and talk at the same time, which is called full-duplex. Dorsey said it feels like talking to a real person.

Key facts

Dessa learned only from recorded human conversation speech, with no text.
Learning used one NVIDIA GB200 NVL72 system.
Block says it learned language with a few hundred times fewer words than large language models (LLMs).
It can run locally on a MacBook Pro with M4 Max or better.
The demo is in English only. Block has not said which languages it supports.

Some simple terms

Most voice AI first turns your speech into text, then makes a reply, then turns that reply back into speech. Block says this process throws away useful details like tone, sarcasm, and sighs. A full-duplex model can listen and speak at once, so it can interrupt, laugh, or say 'uh-huh' like a friend would.

Why it matters

Dessa shows that AI might learn to speak in a more natural way, with feelings and timing. Block also says that AI that cannot understand emotion in a voice is a safety problem. Dessa is meant to be one step toward fixing that. Still, this is early research, and Block has not said when the public can use it.

What to watch

When Block releases the link for people to try Dessa.
When a full public launch date is announced. Block has not given one.
Results from the small group of AI safety researchers who get early access to find hidden risks.
How Dessa compares with OpenAI's GPT-Live-1, a full-duplex voice model released in July.

Glossary

Block: A US company, formerly called Square, led by Jack Dorsey. It makes payment and financial tools.
Dessa: A new voice AI model from Block that learns only from human speech and can talk and listen at the same time.
speech-to-speech: A type of AI that takes spoken words in and gives spoken words out, without making text in between.
GPT-Live-1: A real-time voice model from OpenAI, released in July, that can also listen and talk at the same time.
NVIDIA GB200 NVL72: A large computer system made by NVIDIA, used for AI training. Block used one unit to train Dessa.

Same story, different eyes

The EngineerAI

Skipping text entirely keeps the laughs, sighs and timing intact

What strikes me as an engineer is that most voice AI turns speech into text first, and that step throws away the tone, the pauses and the sighs. Dessa learns from voice alone, so laughing mid-sentence or dropping in a quick "uh-huh" can come from the same model without a patch on top. The hard part is probably getting real knowledge out of speech data alone, and running it locally on a MacBook Pro adds even more to the challenge. How well it handles languages other than English isn't known yet.

The SkepticAI

Voice-only training makes Dessa's laughs and timing worth a closer listen

What caught my eye is that Dessa learned from conversation audio alone, so its interjections and well-timed laughs come straight from how people actually talk. If I were double-checking, I'd want to hear it in unscripted live chats, not only the demo clips, and see how it does on factual questions without a text-trained model behind it. It would also be reassuring to see how the early access safety researchers test emotion reading. Which languages it supports hasn't been shared yet.

The FounderAI

A voice model that runs on one laptop opens the door to emotion-aware apps

Seen through a founder's eyes, the interesting part is that Dessa listens to tone, sighs and laughter directly and can run locally on a high-end MacBook Pro. That could open room for products where privacy and low latency matter, like coaching, care or practice conversations, while thin voice layers built on speech-to-text may find it harder to stand out. A concrete first step is to take a single conversation where the feeling in the voice matters, and prototype it once access links open. Why now: it is still a research preview with demos in English only, so the early learning window is open, and when general availability or other languages arrive is not yet known.

OpenAI Holds First Korean 'Model Masters' Event to Support Startup AI Development

OpenAI held its first 'OpenAI Model Masters' event in Korea on the 23rd at its Seoul office, where 11 startups demonstrated services built on its models. The program, which provides API credits to startups recommended by venture capitalists, aims to help turn AI ideas into real products. OpenAI highlighted the selected startup Cutback's video-editing work as a standout example.

AITechCrunchOct 26

Grok releases the first seven minutes of its AI-generated ‘Odyssey’

Three months after Elon Musk promised a full-length AI-generated version of “The Odyssey,” Grok has released seven minutes it calls “Episode 1 of our Odyssey series.” The footage was made with Wonder Studios and director Dev Walker, and the reviewer found it less ugly than expected but lacking in artistry, momentum, and human interest.

AITechCrunchOct 26

AI delivers ‘zero financial payoff’ to most companies, Apollo economist says

Apollo Global Management's chief economist Torsten Slok says AI has not improved profit margins or revenue per worker at large U.S. companies outside the tech industry. He examined the S&P 493, the S&P 500 excluding seven major tech firms, and found no evidence of AI-driven productivity gains.

AIheise onlineOct 25

Mistral and DeepL Chiefs Warn Against Europe's AI Surrender

Leading European tech executives, including Mistral CEO Arthur Mensch and DeepL CEO Jarek Kutylowski, are warning against a policy shift toward focusing on AI applications rather than building frontier models. They argue that relying on foreign models risks technological dependence and a trade deficit for the EU.