Dessa

About

Full-duplex conversation model trained purely on human speech

Dessa is a full-duplex speech-to-speech model built by an AI research lab at Block led by Ragavan Thurairatnam (founded by former members of the Toronto deep learning startup Dessa, reporting to Jack Dorsey). Unlike typical voice pipelines that transcribe speech to text, run an LLM, and synthesize audio back out, Dessa was trained from scratch on raw human conversation audio with zero text data and no pretrained models, on a single NVIDIA GB200 NVL72. Because it operates directly on audio in and audio out, it hears and produces vocal emotion, laughter, sighs, sarcasm, pauses, and inflection, and it handles natural turn-taking, interruptions, and backchanneling ("yeah!", "got it!") while listening and speaking at the same time. Inference runs faster than real time and fits locally on a MacBook Pro M4 Max or consumer GPUs. Currently in research preview with early access open to AI safety researchers (contact ragavan+dessa@block.xyz) and a public waitlist on the site.

Features

• True speech-to-speech architecture trained from scratch on audio alone (no text or pretrained LLMs)
• Full-duplex conversation: listens and speaks simultaneously with natural interruptions and backchanneling
• Hears and expresses vocal emotion, laughter, sarcasm, sighs, and cadence
• Faster-than-real-time inference that runs locally on a MacBook Pro M4 Max or better
• Trained on a single NVIDIA GB200 NVL72 by Block's Dessa research lab