Spoken language identification for 12 Nigerian languages (DSN 2026)

Submission of team "godzilla" to the DSN 2026 Find the Language I Am Speaking competition.

A standard HubertForSequenceClassification model (loads with AutoModelForAudioClassification, no custom code). It predicts one of 13 labels: yor ibo hau ego anw efi ibb idu igl nup tiv urh other.

Approach

AfriHuBERT (all 12 transformer layers) with a weighted layer sum and mean pooling, fine-tuned for 8,000 steps on random 3-6 s crops with speaker- and source-balanced sampling (external sources capped at half of a class) and audio augmentation (synthetic noise, reverberation, telephone band-limiting and companding, speed and equalisation changes). Training labels were cleaned (clearly English segments relabelled as other). Per-class decision offsets, tuned on speaker-held-out cross-validation, are added to the classifier bias. Inference uses the full clip.

Base model

Training data

  • Competition training and validation data, Data-Science-Nigeria/dsn-2026-lid-train, which combines afrispeech_ipa (CC-BY-4.0), hypa_voices (Apache-2.0) and omnilingual_asr (CC-BY-4.0); see its ATTRIBUTION.md. No other use of those corpora.

External datasets (openly licensed; per-dataset licence shown):

Usage

from transformers import AutoFeatureExtractor, AutoModelForAudioClassification
fe = AutoFeatureExtractor.from_pretrained('godzilla1215/dsn2026-lid-d4')
model = AutoModelForAudioClassification.from_pretrained('godzilla1215/dsn2026-lid-d4')
inputs = fe(waveform_16khz_mono, sampling_rate=16000, return_tensors="pt")
label = model.config.id2label[model(**inputs).logits.argmax(-1).item()]
Downloads last month
25
Safetensors
Model size
94.6M params
Tensor type
F16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for godzilla1215/dsn2026-lid-d4

Finetuned
(32)
this model