Block has introduced Dessa, a new speech model designed for full-duplex conversation — it can listen and speak at the same time. The company says Dessa was trained entirely on speech, can convey pauses and emotion, and can run locally on a MacBook M4 Max, a combination that could change the privacy and latency discussion for voice assistants.

Why full-duplex matters

Most voice assistants work in turns: you speak, they respond, and interruptions break the flow. A full-duplex model that listens while it speaks is closer to a real conversation, where people overlap, correct themselves and change direction mid-sentence. For businesses testing voice assistants with customers, that naturalness could make automated calls feel less robotic — but it also raises the bar for testing interruption handling, accuracy and accessibility.

The authenticity question

Dessa arrives in a week when authenticity is on the business agenda in other ways too. Pennsylvania has signed a law requiring disclosures on significantly AI-generated adverts, and the BBC's Blue Planet III is keeping imperfect footage to help viewers recognise authentic filming. A natural-sounding voice is not evidence of a reliable service, and callers should be clearly told when they are speaking to AI, with an easy route to a person.

What to watch

These are company claims, not independent benchmark results, and Dessa will need to prove itself against real customer tasks before businesses rely on it. The interesting test is not whether it sounds human, but whether it handles the messy reality of a phone call — background noise, hesitation, and a customer who changes their mind halfway through a sentence — without losing the thread.