DSN 2026 spoken language ID (Whisper-base encoder)

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

Variant

Same weights as the calibrated model, with 50% of the per-class calibration offsets removed.

Encoder of openai/whisper-base, fine-tuned end to end with WhisperForAudioClassification (weighted layer sum over encoder layers).
Input window of 10 s: the processor pads or truncates every clip to its first 10 s, and the encoder keeps the matching first positions of Whisper's position table.
Training: rare-language oversampling (inverse-sqrt class frequency), random crops, waveform augmentation (speed, gain, reverberation, telephone band-limiting with mu-law, coloured noise, babble from other training clips), SpecAugment, label smoothing 0.1.
Per-class logit offsets tuned to maximise language-weighted F1, kept only if they also improve held-out halves of the validation split, and folded into the classifier bias.

Train split only. The validation split was used for epoch selection and calibration.
Source: Data-Science-Nigeria/dsn-2026-lid-train. Its sources keep their own licences; see the dataset's ATTRIBUTION.md and LICENSE.md.

best epoch (0-based): 6
val language-weighted F1 (raw): 0.616
val language-weighted F1 (calibrated, in-sample): 0.6811
calibration held-out gain (2-fold): 0.0369
calibration applied: True
per-class F1: {'yor': 0.862, 'ibo': 0.989, 'hau': 0.987, 'ego': 0.917, 'anw': 0.982, 'efi': 0.211, 'ibb': 0.69, 'idu': 0.943, 'igl': 0.215, 'nup': 0.0, 'tiv': 0.58, 'urh': 0.988, 'other': 0.892}

Nupe has no validation clips, so its calibration offset is left at zero.
Training data is mostly read speech, including religious recordings, so conversational speech may score lower.
Licence of this repo: MIT, which is the licence of the base Whisper model. The training data's licences are separate.