Mustyyyy/dsn-lid-afrihubert

Nigerian spoken language identification (afrihubert)

Classifies a 16 kHz mono speech clip into one of 13 labels: yor ibo hau ego anw efi ibb idu igl nup tiv urh other. Submission to the DSN 2026 language ID challenge.

Approach

• Fine-tuned ajesujoba/AfriHuBERT with the standard AutoModelForAudioClassification head (mean pooling, learned weighted sum of all layers).
• Training on random 8 s crops (short clips tiled), evaluation on full clips.
• Class imbalance: sampling with weights count^-0.5, loss weighted by the competition's class shares, label smoothing 0.05.
• Augmentation: additive noise (5 to 30 dB SNR), phone-band simulation (8 kHz round trip), random gain.
• Per-class offsets tuned on validation and folded into the classifier bias.
• Validation competition F1: 0.5953. Weights stored in fp32.

Training data

• Data-Science-Nigeria/dsn-2026-lid-train (train split) only. See that dataset's ATTRIBUTION.md for its sources and licenses.
• External datasets: none used for fine-tuning. The base model ajesujoba/AfriHuBERT was pretrained by its authors on their own data (see its model card).

Base model

Alabi, Liu, Klakow, Yamagishi. AfriHuBERT: A self-supervised speech representation model for African languages. Interspeech 2025. arXiv:2409.20201