dsn2026-lid-mms

Spoken language identification for 12 Nigerian languages plus an other class, built for the DSN 2026 competition Find the Language I Am Speaking.

Approach: the first 12 transformer layers of facebook/mms-300m with a weighted layer sum and a small classification head. The CNN feature encoder was frozen. Trained on random 5 second crops with class balanced sampling, noise, volume, clipping and 8 kHz phone style augmentation, label smoothing, mixed precision, 8 epochs. Weights are stored in fp16.

Training data: only the provided Data-Science-Nigeria/dsn-2026-lid-train dataset (train splits). No external datasets were used. None of the prohibited corpora were used.

Labels: yor ibo hau ego anw efi ibb idu igl nup tiv urh other