Block ships Dessa, a voice model trained without any text

Block, Jack Dorsey’s payments company, unveiled Dessa on 23 October: a full-duplex speech-to-speech model it says was trained from scratch on nothing but human conversation audio. No text, no transcripts, and no pretrained language model sit underneath it. Block calls it the first model you can hold a realistic live conversation with that learned purely from speech.

The whole thing was trained on a single Nvidia GB200 NVL72, and Block says inference runs faster than real time on a MacBook Pro M4 Max or better, with no cloud required. Because it skips text entirely, Dessa hears and reproduces emotion directly: sarcasm, laughter, a frustrated sigh, a voice cracking with sadness. It also knows when to cut in and when to just backchannel with a “yeah” or “got it”, the small sounds that make a text-pipeline voice bot feel robotic.

Dessa is a research preview, not a product. There’s a public waitlist, but Block is only granting early access to AI safety researchers for now, and has not said when, or whether, it will open further.

Why it matters: almost every voice agent today is a pipeline bolted together from speech-to-text, an LLM, and text-to-speech, which is where the stilted timing and lost emotion come from. A model that learns conversational timing and affect straight from audio, and runs locally on consumer hardware, is a different architecture for builders to watch, even if nothing here is production-ready yet.

The caveat: every claim above is Block’s own. There are no independent benchmarks, no public weights, and no committed release date beyond the research preview.