Alexdata26/dsn-2026-lid-low-lr

DSN 2026 Nigerian Language Identification

A 13-class audio-classification model fine-tuned from facebook/wav2vec2-base using the official competition training split.

Labels

Exact output order, indexed from 0 to 12:

yor ibo hau ego anw efi ibb idu igl nup tiv urh other

ego means Eggon. other represents speech outside the 12 target languages.

Model and preprocessing

• Standard Hugging Face audio-classification architecture.
• Loads using AutoModelForAudioClassification without custom code.
• 94,571,917 parameters.
• Stored weights: fp32 safetensors.
• Input: 16 kHz mono audio.
• Feature extractor: normalization enabled, right-side zero padding, return_attention_mask=False.

No custom cropping, window aggregation, or logit calibration is embedded in the repository.

Training data

Official dataset: https://huggingface.co/datasets/Data-Science-Nigeria/dsn-2026-lid-train

Pinned dataset revision: 87cf6d4b1c7599f97f44b6e6aade86438369a2b1

• Training: 24,774 clips.
• Validation: 4,352 clips.
• Only the training split was used for gradient updates.
• Validation was used for evaluation and checkpoint selection.
• No additional external training dataset was used.
• Pretrained initialization: facebook/wav2vec2-base.

The upstream corpora listed in the official dataset attribution were not downloaded separately or used directly as additional training data.

Training and checkpoint selection

Original fine-tuning used batch size 4, AdamW learning rate 3e-5, per-example sampling weights proportional to 1/sqrt(class_count), and random crops up to 8 seconds.

The selected checkpoint is an 800-update continuation from epoch 1, using a fresh optimizer at learning rate 5e-6, the original sampler, batch size 4, and random crops up to 8 seconds.

The convolutional feature encoder was frozen during continuation. Training used AMP fp16; the export stores fp32 parameters.

This checkpoint comes from the recovery run completed on 2026-10-21.

Local validation

Evaluation used full clips, batch size 1, and AMP fp16.

Model	Represented-class diagnostic
Original epoch 1	0.5684942873
Selected low-LR step 800	0.5806749508

The diagnostic uses the stated fixed class shares, normalized over classes represented in validation.

This is not the official competition score. Nupe has no validation examples, so Nupe recall and F1 are unmeasured. Do not compare this diagnostic directly with the official 0.60 ranking threshold.

Predictions into Nupe still count as errors for the true represented classes. No Nupe-specific logit calibration was applied.

The hidden evaluator's exact preprocessing, batching, and precision were not available in the supplied starting kit.

Limitations

• Efik and Ibibio have zero F1 in the documented local evaluation.
• Igala and Tiv remain weak.
• Several languages have very few known training speakers.
• Nupe has one known training speaker and no validation support.
• Local results do not establish hidden-test performance on unseen speakers, noisy recordings, or phone-quality audio.
• Gains in some classes accompany regressions in others, including Hausa and Annang.
• No winning result or ranking eligibility is claimed.
• This model is not suitable for high-stakes decisions about speakers.

Checks performed

This selected export passed:

• Standard Hugging Face offline CPU functional loading.
• The supplied starting-kit model-folder file/header checker.

These checks do not certify speed, peak RAM/GPU memory, total runtime, or compatibility with every Transformers version.

Local environment: Transformers 5.16.1 and PyTorch 2.11.0+cu128.

Attribution and licensing

The official dataset declares mixed per-source licensing. Its supplied LICENSE.md is preserved unchanged in this repository. The supplied attribution notice is reproduced below.

Attribution

This dataset combines the sources below. Keep their attribution and license terms.

• afrispeech_ipa: CC-BY-4.0 (stated; upstream audio terms not verified)
• hypa_voices: Apache-2.0
• omnilingual_asr: CC-BY-4.0

Source documentation

• Dataset: https://huggingface.co/datasets/Data-Science-Nigeria/dsn-2026-lid-train
• Pinned attribution: https://huggingface.co/datasets/Data-Science-Nigeria/dsn-2026-lid-train/blob/87cf6d4b1c7599f97f44b6e6aade86438369a2b1/ATTRIBUTION.md
• Pinned license notice: https://huggingface.co/datasets/Data-Science-Nigeria/dsn-2026-lid-train/blob/87cf6d4b1c7599f97f44b6e6aade86438369a2b1/LICENSE.md
• Pretrained model: https://huggingface.co/facebook/wav2vec2-base

The dataset card documents resampling, segmentation, label mapping, deduplication, and speaker-based splitting. This model adds supervised fine-tuning and continued fine-tuning. Source audio is not redistributed.

The pretrained model card declares Apache-2.0. This does not establish a blanket Apache-2.0 license for all training-data rights or derivatives.

The dataset documentation supplies this citation:

@misc{omnilingualasr2025,
    title={{Omnilingual ASR}: Open-Source Multilingual Speech Recognition for 1600+ Languages},
    author={{Omnilingual ASR Team} and others},
    year={2025},
    eprint={2511.09690},
    archivePrefix={arXiv},
    primaryClass={cs.CL},
    url={https://arxiv.org/abs/2511.09690}
}

The supplied source notices state that some upstream audio terms are unverified. No independent verification or blanket licensing assurance is claimed here.