Block’s new voice model sounds eerily human
Any company can demo an impressive voice model these days. Block saying theirs runs locally on a MacBook M4 Max is the detail that actually matters. Local inference changes the economics of voice AI. No per-minute cloud fees, no network dependency, and for a payments company, no customer audio leaving the device. That last point is probably the real motivation. Block sits on sensitive financial data, and regulators are already nervous about how AI handles it.
I'd want to know how the model sounds in practice, because "eerily human" in a demo and "eerily human" in a noisy kitchen are different claims. The M4 Max is also a high-end chip, so this isn't the same as a voice model running on a phone. Still, the direction is clear. Voice interfaces have been waiting for a model that's good enough and cheap enough to run locally, and this looks like a step toward that.