Jack Dorsey, the founder of Twitter now known as X, has returned with a new artificial intelligence model named Dessa that operates without any underlying large language model. Traditional voice assistants like ChatGPT and Gemini typically convert spoken input into text prior to formulating a response, a process that can strip away nuanced details such as sarcasm, emotional inflection, laughter, and natural speech pauses.
Dessa adopts an alternative approach created by a small research laboratory within Dorsey's firm, Block. The model acquires language capabilities by directly listening to human speech instead of reading written text. Developers report that the system supports real-time, natural conversations and executes locally on an M4 Max MacBook Pro or superior hardware configuration, entirely independent of an internet connection. Audio samples shared across the project's website and X have generated significant public interest.
Unlike standard voice assistants that require a user to finish speaking before they reply, Dessa incorporates full-duplex speech-to-speech technology, allowing it to listen and talk simultaneously. The technology enables the AI to interrupt naturally, offer brief acknowledgements like yeah or got it, and laugh mid-sentence during humorous moments.
Developers state that the model detects emotional cues, sarcasm, and tonal shifts that standard text conversion often misses. Trained utilizing a single NVIDIA GB200 NVL72 system, Dessa represents a fresh direction for artificial intelligence investigation into how models might absorb language and factual information purely through sound.
Responding to inquiries on X, Dorsey confirmed that the system utilizes direct speech-to-speech processing without any LLM and verified that it functions locally. Creators are currently granting early access to a select group of safety researchers to evaluate potential risks, positioning the project as an exploratory look at communication models built on listening rather than reading.