Block’s voice model claims to sound human, but it’s not talking to us

The story

Block released a voice model called Dessa that runs locally and mimics human speech with pauses, emotions, and cadence, but it's not publicly available yet.

Source: techcrunch.com

Our reading

Block’s new voice model, Dessa, is supposed to sound so human that it can mimic tiredness, laughter, and even emotional breakdowns. The company says it’s the most authentic full-duplex speech model ever built, which means it can talk and listen at the same time. But the real kicker is that it’s trained entirely on human speech, not text. That’s a weird detail. Most voice models rely on text-to-speech pipelines, which convert spoken words into written text before generating a response. Dessa skips that step, which is supposed to make it feel more natural. But if it’s not using text, how does it understand context? And why would that make it sound more human? Maybe it’s just better at pretending. Either way, the model is still locked behind a wall, and Block is asking safety researchers to test it. Which is a bit like asking your friend to check if their new AI is trying to trick you, but they won’t let you see it.