Speech brain-computer interfaces face challenges scaling across individuals with different brain organization. Using minimally invasive recordings from 25 patients, the authors developed transfer learning methods that enable robust speech decoding even with incomplete brain coverage.