
ⓘ Andres Siimon on Unsplash
Block has unveiled Dessa, a new speech-to-speech AI model trained entirely on human conversations without written text or pretrained language models. The system supports natural, real-time conversations with interruptions and emotional cues, and reportedly runs locally on a MacBook Pro equipped with Apple's M4 Max chip.
Jack Dorsey’s Block, the company behind Square and Cash App, has unveiled Dessa, a new artificial intelligence model designed to hold conversations almost like a human. Unlike conventional voice assistants that rely on text processing, Dessa has reportedly learned everything it knows simply by listening to human conversations.
Dessa is described as a full-duplex speech-to-speech model. In simpler terms, it can listen and speak simultaneously, allowing it to interrupt conversations, acknowledge what someone is saying and react without waiting for the other person to finish talking.
According to Block’s official announcement, Dessa was trained entirely from scratch using recordings of human conversations. It reportedly uses neither written text nor pretrained language models, meaning its understanding of language, facts and conversational behaviour comes exclusively from audio.
The company claims this approach allows Dessa to recognise nuances that traditional speech-to-text pipelines might overlook, including sarcasm, hesitation, laughter and changes in emotional tone. In several demonstrations published on its website, the model laughs mid-sentence, raises its voice when asked to speak louder and even finishes a sentence when someone struggles to remember a movie title.
One particularly interesting demonstration features Dessa holding a conversation with OpenAI’s GPT-Live-1 voice model. Block says these recordings capture live interactions without anyone directing the model’s responses. However, they are first-party demonstrations and should be treated as such.
There’s also an interesting hardware angle. Block claims Dessa was trained using just one Nvidia GB200 NVL72 system, which houses 72 Blackwell GPUs. The resulting model can reportedly run locally on a MacBook Pro equipped with Apple’s M4 Max chip or better. Block also claims inference runs faster than real time, allowing conversations without noticeable pauses.
However, Dessa isn’t publicly available just yet. Block is inviting a limited number of AI safety researchers to evaluate the model, with a waitlist available for everyone else. The company has also not disclosed its parameter count, model weights or standardised benchmark results.
Anil Ganti - Senior Tech Writer - 3147 articles published on Notebookcheck since 2019
I've been an avid PC gamer since the age of 8. My passion for gaming eventually pushed me towards general tech, and I got my first writing gig at the age of 19. I have a degree in mechanical engineering and have worked in the manufacturing industry and a few other publications like Wccftech before joining Notebookcheck in November 2019. I cover a variety of topics including smartphones, gaming, and computer hardware.
Anil Ganti, 2026-10-24 (Update: 2026-10-24)

