ogbebormichael/dsn-lid-mhubert-v1

DSN 2026 spoken language identification (v1)

Approach: utter-project/mHuBERT-147 (HuBERT base architecture, CC-BY-NC-SA-4.0), keeping the first 9 of 12 transformer layers, with a mean-pooling classification head over 13 labels (12 Nigerian languages plus other). Fine-tuned with class-balanced sampling, random 3-6 second crops, and augmentation (background noise, speed and pitch change, phone-quality band-limiting, volume change). Weights saved in fp16.

Training data: only the competition-provided dataset Data-Science-Nigeria/dsn-2026-lid-train. External datasets: none. No data from the prohibited corpora was added. Checkpoint: step 3000, validation weighted F1 0.591.