Block previews Dessa, a local voice model with full-duplex conversation
The AMW Read
Dessa meaningfully adds to the conversational model player map through local full-duplex speech, but restricted access and unproven production economics limit its demonstrated impact.
Block previews Dessa, a local voice model with full-duplex conversation
Block has introduced Dessa, a voice model from its research lab that can speak and listen simultaneously. The company says it produces laughter, pauses and crying, with emotion built into the model. Block says Dessa was trained entirely on human speech using a single NVIDIA GPU rack. Its knowledge comes from conversational training data rather than a paired large language model. Dessa can run on consumer hardware, including a MacBook M4 Max. It is not publicly available; Block is inviting safety researchers to test it.
Dessa adds a locally runnable conversational model to a market where speech interfaces can discard vocal information when converting speech into text. Block argues that this conversion loses emotional cues. The market question is whether a model trained directly on speech can combine expressive interaction with useful conversational knowledge. Local execution also makes deployment economics relevant: the article identifies potential cost savings for Block, whose Square unit already uses voice AI for ordering. Neither demos nor hardware compatibility establish those savings in production.
Builders and investors should distinguish expressive demonstrations from readiness for customer workflows. A concrete evaluation would test simultaneous listening and speaking, conversational accuracy and operating cost on the stated consumer hardware. The reported training setup alone does not establish training cost or efficiency against competing models. Until testing supports broader use, Dessa is a research preview with a possible internal distribution path through Block, rather than evidence of a deployed ordering system.
