Block Releases Dessa Voice Model on 1 Nvidia GPU as It Pushes More Human AI
Updated
Updated · TechCrunch · Oct 24
Block Releases Dessa Voice Model on 1 Nvidia GPU as It Pushes More Human AI
3 articles · Updated · TechCrunch · Oct 24
Summary
Block said Dessa can run locally on a consumer machine such as a MacBook M4 Max, with the voice model trained on human speech using a single Nvidia GPU.
The company argues that avoiding a text-first pipeline preserves emotional cues that current voice AI often loses, letting Dessa speak with laughs, pauses, cries and other more human-sounding expressions.
Dessa is full-duplex—able to listen and talk at the same time—and Block showed demos with different voices, cadences and moods, including a tired-sounding bot and an emotional exchange.
Block has not opened public access and is instead inviting safety researchers to probe whether a more lifelike voice agent could mislead people, a risk with obvious relevance for finance and customer-service use cases.