GPU

An experiment to format dictation in under 100 milliseconds

A 60M-parameter T5, distilled from a larger model, that cleans up raw speech on-device — split across the Neural Engine and the GPU.