Block Unveils Dessa: A Voice Model That Learned to Speak From Human Speech Alone

Block's in-house AI research lab unveiled Dessa on October 23, a full-duplex speech model that takes a different route from mainstream voice assistants: no text in the middle, audio in and audio out.

There are three highlights. First, zero-text training. Dessa learned language from scratch by listening only to real human conversation, with no text data and no pretrained model; its lead calls it a "GPT-3 moment" for speech. Second, it feels like a phone call. Full-duplex means it listens while it speaks, so it can interrupt, backchannel, and catch laughter, sarcasm, and a cracking voice, with inference faster than real time. Third, it runs locally. The model can run on a MacBook Pro M4 Max or better, and training used a single NVIDIA GB200 NVL72 rack.

One caveat: Dessa is currently only a research preview. Early access goes to a small group of AI safety researchers, with a public waitlist open; Block has published no parameter count or third-party benchmarks, so the demo results still await independent verification.

Source link: Read the original, this article is a summarized digest for reference only.