Obtain consent terms and freeze speaker-disjoint splits

Obtain consent terms and freeze speaker-disjoint splits

Identify a consented Swahili speech corpus, record its license terms and permitted uses, and define train, development, and test splits in which no speaker appears twice. Label dialect groups before any model runs. Deliverable: A split manifest with speaker and dialect counts, a license record, and a statement that freezing the splits does not itself reduce word error rate.