What happened?
Block introduced Dessa, a research preview trained directly on speech that can listen and speak at the same time.
Why does it matter?
Voice AI needs to handle interruptions and uncertainty, as well as sound natural. Independent, unscripted testing is the next useful test.
What happened?
Block introduced Dessa on October 23, with weekend coverage following on Saturday. Its launch page describes a model trained directly on human speech that can listen and speak simultaneously, a capability called full duplex. Block says it runs locally on a MacBook Pro with an M4 Max chip or better.
The company is inviting a small group of AI safety researchers to seek out unanticipated risks. The page presents recordings and access applications rather than a generally available product. Its performance descriptions are developer claims.
Source: Block ↗ · Source: Numerama ↗
Why does it matter?
Our analysis: a useful voice interface must do more than produce a convincing voice. It needs to manage interruptions, pauses and uncertainty without making a conversation harder to follow. That is a different design problem from delivering a polished answer after a user has finished speaking.
The next assessment should involve long, unscripted exchanges: noisy rooms, ambiguous statements, changes of topic and ordinary misunderstandings. A system that sounds fluent in a short clip can still be difficult to use when a conversation goes off course.
Emotional expression deserves similarly careful evaluation. A laugh or sympathetic tone can shape a user’s expectations, but it does not establish sound judgment. Independent testing should examine what the system does when the appropriate response is to slow down, ask for clarification or acknowledge a limit.
Read the sources
This article was drafted with AI assistance and checked against the linked sources. Reporting summarizes those sources; sections labeled “Our take” contain editorial analysis. The underlying event date is shown separately from this article’s publication date. Our editorial standards ↗
