Block, the company led by Jack Dorsey, has unveiled Dessa, a full-duplex speech model trained from scratch on nothing but human conversation. Block calls it a research preview and says it is the first AI you can hold a realistic live conversation with that never saw a word of text during training. No text corpora, no pretrained backbone, no fine-tuning on transcripts. The company says Dessa learned to converse simply by listening.
The pitch is that dropping text preserves everything transcription throws away: emotion, pauses, timing and cadence. Dessa speaks and listens at the same time. It laughs mid-sentence when something is funny, drops its voice when a moment turns serious, and gets louder when asked, without anyone teaching it to. Dorsey described the model as big but also really really small, noting that it is small enough to run on a MacBook. That combination of full-duplex real-time conversation and local hardware is what turned heads.
The launch landed hard on developer forums, where the thread drew several hundred points and a mix of excitement and scepticism. Supporters called it the strongest speech-to-speech demo yet and pointed out that real-time voice has been dominated by one or two labs. Sceptics want to know how Dessa handles language it never heard, whether the weights will ever be opened, and how much compute the training run consumed. Block has not published the weights. Early access is limited to AI safety researchers who apply through the company site, and there is no pricing or public release date yet. For founders building voice products, Dessa is a signal that full-duplex, speech-native models are turning into a competitive category rather than a research curiosity.