Omnilingual ASR: Open-Source Multilingual Speech Recognition for 1600+ Languages
Paper • 2511.09690 • Published • 1
dsn2026-lid-s01: spoken language identification for 12 Nigerian languages
Submission to the DSN 2026 competition "Find the Language I Am Speaking". Given a speech clip, the model predicts one of 13 labels:
yor ibo hau ego anw efi ibb idu igl nup tiv urh other.
It is a standard AutoModelForAudioClassification checkpoint (fp16 safetensors), with no custom code.
Team
- Oludare Samuel
- Fagbemiro Victoria
- Sumayah Adegbite
Approach
Base model:
ajesujoba/AfriHuBERT(encoder only, 94.6M parameters), fine-tuned end to end with a weighted-layer-sum pooling head.Training: random 2-10 s crops; class-rebalanced sampling (rare languages drawn more often); label smoothing 0.1.
Augmentation (generated from the training audio itself, no external audio): coloured noise, babble from other training clips, simulated reverberation, telephone band-limiting with mu-law companding and clipping, random gain, and speed perturbation (0.85-1.15x) for speaker robustness.
Calibration: per-class logit offsets tuned on the validation split to maximise the competition's language-weighted F1. They are added to the classifier bias, so inference is a plain argmax.
Export: fp16 weights.
Training data
- Data-Science-Nigeria/dsn-2026-lid-train (train and validation splits).
External datasets (openly licensed; none of them is a corpus listed in the competition's PROHIBITED_SOURCES.txt), used for training only:
- offiongbassey/efik_speech_dataset (CC-BY-4.0): Efik speech, 2 speakers; a random 3 h subset. From "Towards Digital Preservation of Efik: TTS for a Low-Resource African Language" (Interspeech 2026).
- Mozilla Common Voice Scripted Speech 27.0, Ibibio (CC0-1.0): validated and not-yet-reviewed clips (invalidated excluded), at most 100 clips per speaker, 19 speakers, 3.6 h.
The pretrained base model is ajesujoba/AfriHuBERT.
Attribution of the training data sources
The official training data combines these sources, each under its own license:
- afrispeech_ipa: CC-BY-4.0
- hypa_voices: Apache-2.0
- omnilingual_asr: CC-BY-4.0. If you use this portion in research, please cite: Omnilingual ASR Team et al., "Omnilingual ASR: Open-Source Multilingual Speech Recognition for 1600+ Languages", arXiv:2511.09690, 2025.
Validation results (local)
Uniform weight average (model soup) of two AfriHuBERT fine-tunes from the same base: F05 (source-debiased recipe, train+validation) and F07 (plain recipe, train+validation + external Efik and Ibibio). On local validation, the same soup of the two recipes' train-only versions scored weighted F1 0.626 vs 0.608 and 0.594 for the single models.
Limitations
Several languages have very few training speakers (Nupe 1, Igala 2, Ibibio 2), so accuracy on unseen speakers of those languages is uncertain. Annang, Ibibio and Efik are closely related and can be confused.
- Downloads last month
- -