Block unveils Dessa, a voice model trained only on human conversation

What this is about

On October 23, 2026, the AI research lab at Block (Jack Dorsey's company) unveiled Dessa. Dessa is a so-called speech-to-speech model: audio in, audio out, with no detour through text. According to Block, it was trained "from scratch on human conversation only," meaning sound alone, with no text and no pretrained models. Block calls it the "most authentic full-duplex voice model ever built." Project lead Ragavan Thurairatnam called it a "GPT-3 moment for voice" on X. So far, Block has released demo recordings, a public waitlist and early access for a small group of AI safety researchers; launch links are due "soon." All claims come from Block itself and have not been independently verified.

What Dessa is supposed to do

Full-duplex means the model listens and speaks at the same time, like people on the phone. Dessa is said to have learned from real conversations when to chime in and when to give backchannel cues ("yeah!", "got it!", "ooh"). It is supposed to hear and produce emotions, sarcasm, laughter, a frustrated sigh or a breaking voice directly. Inference (generating the responses) reportedly runs faster than real time, so there is no waiting pause, and the model runs locally on a MacBook Pro M4 Max or better.

Block criticizes conventional voice AI: it converts speech to text and throws away a lot of information in the process, such as how someone feels. That leads to stiff, slow conversations.

The demos

Block shows recordings of real conversations in which, according to the company, nobody is steering the model; subtitles were added afterward. Among other things, they describe:

• Dessa laughs mid-sentence when something is funny, rather than after the speaker finishes.
• It gets louder when you ask it to speak up.
• It finishes the sentence when the other person forgets a movie title.
• Given a surprise message, it gasps softly and replies with something that sounds like touched gratitude.
• There is a longer recording with empathetic life advice, as well as a conversation with GPT-Live-1, a third-party model. Both voices are AI-generated; its developer is not affiliated with Block and has not reviewed Dessa.

The Verge describes the conversations as mostly natural in their turn-taking, but occasionally a bit awkward and uncanny. The model initially mistook the speaker for a "Ronaldo" and drifted from a conversation about Christopher Nolan's films to Bryan Singer's Superman films. In the conversation with GPT-Live-1, Dessa comes across as more natural.

Why speech is so hard

Block argues that speech data is far harder to work with than text. A spoken word takes up thousands of times more bytes than a written one, and spoken sentences never repeat exactly: saying the same sentence twice still produces different data. Emphasis changes meaning, as Block shows with the sentence "I didn't say she stole my CUDA code." Then there are accents, pauses, breathing, humor and timing.

Dessa reportedly learned speech from hundreds of times fewer words than large language models (LLMs), and the team says it found a way to train it on a single NVIDIA GB200 NVL72 (a server-rack system). Open research questions, according to Block: How efficiently do such models learn language and facts compared with text? How efficiently do they learn new languages? And what kind of in-context learning (learning from examples within the conversation) can they do?

Motivation, safety and team

The team gives three reasons: curiosity, better understanding between humans and AI, and safety. Babies learn to speak without ever seeing a written word. On safety: AI is gaining power and independence but barely picks up on emotion, for example when someone is in severe pain and the voice AI doesn't notice. Dessa is meant as a step against that. A small group of safety researchers gets early access to find unforeseen risks; the form requires an institutional email address.

The team is a small lab inside Block, founded by former members of the Toronto startup Dessa, which according to Block has been building deep-learning models since 2012 and generative voice models for more than eight years. It reports directly to Jack Dorsey. Thurairatnam founded and sold the startup Dessa.

Our take

If it holds up, a voice model without text pretraining would be a remarkable approach, capturing emotion and timing directly instead of losing them in text. So far, though, there is no paper and there are no weights, no benchmarks, and no details on model size, languages or training data. The "world's first" claim and the local runtime are vendor statements, and the demos are cherry-picked. The claim that the model knows a lot "from raw sound" can't be verified yet either. Whether and when it will be runnable openly is unclear, so for now Dessa is not an open-weights release. The comparison with OpenAI's GPT-Live-1 stands out, though its developer is explicitly not involved.