Prestigeeo/dsn-2026-lid-xlsr

DSN 2026: Nigerian Spoken Language Identification (Whisper-Base 10s Lid-Calibrated)

Model Summary

This model is a fine-tuned openai/whisper-base (WhisperForAudioClassification) built for the DSN 2026: Find the Language I Am Speaking challenge.

• Architecture: WhisperForAudioClassification with learned weighted encoder layer sum (use_weighted_layer_sum=True), classifier_proj_size=256, and 10-second positional encoding window (max_source_positions=500, chunk_length=10, nb_max_frames=1000).
• Parameters: 20,247,573 (~20.2M parameters, well below the 350M limit).
• Precision & Size: Saved in pure torch.float32 (safetensors, ~81 MB total repository size, well below the 1.5 GB limit) to guarantee zero CPU/GPU dtype mismatch (float16 vs float32) during standard AutoModelForAudioClassification evaluation.
• CPU Inference Latency: ~0.028 seconds per clip on CPU (< 0.5s limit; < 3.5 minutes for all 6,659 test clips).

Training Data & Attribution

Trained exclusively on the official competition dataset (Data-Science-Nigeria/dsn-2026-lid-train, 29,126 clips across 13 classes). Zero external audio datasets were used.

We gratefully acknowledge the upstream open speech corpora curated in the competition dataset:

• AfriSpeech-200 / AfriSpeech-IPA (afrispeech_ipa)
• Hypa Voices (hypa_voices)
• OmniLingual ASR (omnilingual_asr)

Training & Calibration Methodology

1. 13-Class Validation Split: Because the official validation split contains 0 clips for Nupe (nup), we hold out 36 Nupe clips from train into a 13-class validation split (4,388 clips) for full 13-class evaluation and calibration.
2. Class-Balanced Sampling & Competition-Weighted Loss: Trained with inverse-frequency WeightedRandomSampler and label-smoothed cross-entropy weighted by competition evaluation weights (0.087 for the 10 low-resource/confuser classes, 0.043 for yor, ibo, hau).
3. Robust Waveform & Spectrogram Augmentations: Speed perturbation (0.9x–1.1x), simulated 8 kHz telephone bandpass, additive background noise (12–28 dB SNR), random gain, and GPU SpecAugment.
4. Post-Hoc Logit Offset Calibration: Per-class logit offsets are optimized on validation logits to directly maximize competition Language-Weighted F1 and folded directly into model.classifier.bias.

Validation Results (Epoch 4)

• 13-Class Language-Weighted F1 (Calibrated): 0.7020 (Uncalibrated: 0.6122)

Language Code	Competition Weight	Validation F1
yor	0.043	0.8333
ibo	0.043	0.9588
hau	0.043	0.9615
ego	0.087	0.9005
anw	0.087	0.8883
efi	0.087	0.2254
ibb	0.087	0.3732
idu	0.087	0.8650
igl	0.087	0.2745
nup	0.087	0.9091
tiv	0.087	0.4223
urh	0.087	0.9782
other	0.087	0.8638