Block unveils Dessa, a speech model that runs locally on a MacBook

Block, the company behind Square and Cash App, introduced Dessa, a model that processes audio directly instead of transcribing it to text first. It is designed to recognize emotion, sarcasm and pauses in real time. It runs locally on a MacBook Pro M4 Max but is not publicly available yet.

Why it matters: Native speech models that skip the text step can keep tone and timing that transcription discards, which matters for voice agents. Running on a laptop fits the recent push toward small local voice models, such as NVIDIA's Nemotron Bolt, and a fintech company building one suggests customer-facing voice is a competitive area.